Mention detection is an important preprocessing step for annotation and interpretation in applications such as NER and coreference resolution, but few stand-alone neural models have been proposed able to handle the full range of mentions. In this work, we propose and compare three neural network-based approaches to mention detection. The first approach is based on the mention detection part of a state of the art coreference resolution system; the second uses ELMO embeddings together with a bidirectional LSTM and a biaffine classifier; the third approach uses the recently introduced BERT model. Our best model (using a biaffine classifier) achieves gains of up to 1.8 percentage points on mention recall when compared with a strong baseline in a HIGH RECALL coreference annotation setting. The same model achieves improvements of up to 5.3 and 6.2 p.p. when compared with the best-reported mention detection F1 on the CONLL and CRAC coreference data sets respectively in a HIGH F1 annotation setting. We then evaluate our models for coreference resolution by using mentions predicted by our best model in start-of-the-art coreference systems. The enhanced model achieved absolute improvements of up to 1.7 and 0.7 p.p. when compared with our strong baseline systems (pipeline system and end-to-end system) respectively. For nested NER, the evaluation of our model on the GENIA corpora shows that our model matches or outperforms state-of-the-art models despite not being specifically designed for this task.
Anaphora resolution (coreference) systems designed for the CONLL 2012 dataset typically cannot handle key aspects of the full anaphora resolution task such as the identification of singletons and of certain types of non-referring expressions (e.g., expletives), as these aspects are not annotated in that corpus. However, the recently released dataset for the CRAC 2018 Shared Task can now be used for that purpose. In this paper, we introduce an architecture to simultaneously identify non-referring expressions (including expletives, predicative s, and other types) and build coreference chains, including singletons. Our cluster-ranking system uses an attention mechanism to determine the relative importance of the mentions in the same cluster. Additional classifiers are used to identify singletons and non-referring markables. Our contributions are as follows. First all, we report the first result on the CRAC data using system mentions; our result is 5.8% better than the shared task baseline system, which used gold mentions. Second, we demonstrate that the availability of singleton clusters and non-referring expressions can lead to substantially improved performance on non-singleton clusters as well. Third, we show that despite our model not being designed specifically for the CONLL data, it achieves a score equivalent to that of the state-of-the-art system by Kantor and Globerson (2019) on that dataset.
This article introduces Mandarinograd, a corpus of Winograd Schemas in Mandarin Chinese. Winograd Schemas are particularly challenging anaphora resolution problems, designed to involve common sense reasoning and to limit the biases and artefacts commonly found in natural language understanding datasets. Mandarinograd contains the schemas in their traditional form, but also as natural language inference instances (ENTAILMENT or NO ENTAILMENT pairs) as well as in their fully disambiguated candidate forms. These two alternative representations are often used by modern solvers but existing datasets present automatically converted items that sometimes contain syntactic or semantic anomalies. We detail the difficulties faced when building this corpus and explain how weavoided the anomalies just mentioned. We also show that Mandarinograd is resistant to a statistical method based on a measure of word association.
Coreference resolution (CR) aims to find all spans of a text that refer to the same entity. The F1-Scores on these task have been greatly improved by new developed End2End-approaches and transformer networks. The inclusion of CR as a pre-processing step is expected to lead to improvements in downstream tasks. The paper examines this effect with respect to word embeddings. That is, we analyze the effects of CR on six different embedding methods and evaluate them in the context of seven lexical-semantic evaluation tasks and instantiation/hypernymy detection. Especially in the last tasks we hoped for a significant increase in performance. We show that all word embedding approaches do not benefit significantly from pronoun substitution. The measurable improvements are only marginal (around 0.5% in most test cases). We explain this result with the loss of contextual information, reduction of the relative occurrence of rare words and the lack of pronouns to be replaced.
Ellipsis resolution has been identified as an important step to improve the accuracy of mainstream Natural Language Processing (NLP) tasks such as information retrieval, event extraction, dialog systems, etc. Previous computational work on ellipsis resolution has focused on one type of ellipsis, namely Verb Phrase Ellipsis (VPE) and a few other related phenomenon. We extend the study of ellipsis by presenting the No(oun)El(lipsis) corpus - an annotated corpus for noun ellipsis and closely related phenomenon using the first hundred movies of Cornell Movie Dialogs Dataset. The annotations are carried out in a standoff annotation scheme that encodes the position of the licensor, the antecedent boundary, and Part-of-Speech (POS) tags of the licensor and antecedent modifier. Our corpus has 946 instances of exophoric and endophoric noun ellipsis, making it the biggest resource of noun ellipsis in English, to the best of our knowledge. We present a statistical study of our corpus with novel insights on the distribution of noun ellipsis, its licensors and antecedents. Finally, we perform the tasks of detection and resolution of noun ellipsis with different classifiers trained on our corpus and report baseline results.
We present in this work a new dataset of coreference annotations for works of literature in English, covering 29,103 mentions in 210,532 tokens from 100 works of fiction published between 1719 and 1922. This dataset differs from previous coreference corpora in containing documents whose average length (2,105.3 words) is four times longer than other benchmark datasets (463.7 for OntoNotes), and contains examples of difficult coreference problems common in literature. This dataset allows for an evaluation of cross-domain performance for the task of coreference resolution, and analysis into the characteristics of long-distance within-document coreference.
Dramatic texts are a highly structured literary text type. Their quantitative analysis so far has relied on analysing structural properties (e.g., in the form of networks). Resolving coreferences is crucial for an analysis of the content of the character speech, but developing automatic coreference resolution (CR) systems depends on the existence of annotated corpora. In this paper, we present an annotated corpus of German dramatic texts, a preliminary analysis of the corpus as well as some baseline experiments on automatic CR. The analysis shows that with respect to the reference structure, dramatic texts are very different from news texts, but more similar to other dialogical text types such as interviews. Baseline experiments show a performance of 28.8 CoNLL score achieved by the rule-based CR system CorZu. In the future, we plan to integrate the (partial) information given in the dramatis personae into the CR model.
This paper investigates the problem of entity resolution for email conversations and presents a seed annotated corpus of email threads labeled with entity coreference chains. Characteristics of email threads concerning reference resolution are first discussed, and then the creation of the corpus and annotation steps are explained. Finally, performance of the current state-of-the-art deep learning models on the seed corpus is evaluated and qualitative error analysis on the predictions obtained is presented.
Humans do not make inferences over texts, but over models of what texts are about. When annotators are asked to annotate coreferent spans of text, it is therefore a somewhat unnatural task. This paper presents an alternative in which we preprocess documents, linking entities to a knowledge base, and turn the coreference annotation task – in our case limited to pronouns – into an annotation task where annotators are asked to assign pronouns to entities. Model-based annotation is shown to lead to faster annotation and higher inter-annotator agreement, and we argue that it also opens up an alternative approach to coreference resolution. We present two new coreference benchmark datasets, for English Wikipedia and English teacher-student dialogues, and evaluate state-of-the-art coreference resolvers on them.
Coreference resolution aims at identifying and grouping all mentions referring to the same entity. In French, most systems run different setups, making their comparison difficult. In this paper, we present an extensive comparison of several coreference resolution systems for French. The systems have been trained on two corpora (ANCOR for spoken language and Democrat for written language) annotated with coreference chains, and augmented with syntactic and semantic information. The models are compared with different configurations (e.g. with and without singletons). In addition, we evaluate mention detection and coreference resolution apart. We present a full-stack model that outperforms other approaches. This model allows us to study the impact of mention detection errors on coreference resolution. Our analysis shows that mention detection can be improved by focusing on boundary identification while advances in the pronoun-noun relation detection can help the coreference task. Another contribution of this work is the first end-to-end neural French coreference resolution model trained on Democrat (written texts), which compares to the state-of-the-art systems for oral French.
In languages like Arabic, Chinese, Italian, Japanese, Korean, Portuguese, Spanish, and many others, predicate arguments in certain syntactic positions are not realized instead of being realized as overt pronouns, and are thus called zero- or null-pronouns. Identifying and resolving such omitted arguments is crucial to machine translation, information extraction and other NLP tasks, but depends heavily on semantic coherence and lexical relationships. We propose a BERT-based cross-lingual model for zero pronoun resolution, and evaluate it on the Arabic and Chinese portions of OntoNotes 5.0. As far as we know, ours is the first neural model of zero-pronoun resolution for Arabic; and our model also outperforms the state-of-the-art for Chinese. In the paper we also evaluate BERT feature extraction and fine-tune models on the task, and compare them with our model. We also report on an investigation of BERT layers indicating which layer encodes the most suitable representation for the task.
Non-nominal co-reference is much less studied than nominal coreference, partly because of the lack of annotated corpora. We explore the possibility to exploit parallel multilingual corpora as a means of cheap supervision for the classification of three different readings of the English pronoun ‘it’: entity, event or pleonastic, from their translation in several languages. We found that the ‘event’ reading is not very frequent, but can be easily predicted provided that the construction used to translate the ‘it’ example is a pronoun as well. These cases, nevertheless, are not enough to generalize to other types of non-nominal reference.
This paper proposes a new dataset, MuDoCo, composed of authored dialogs between a fictional user and a system who are given tasks to perform within six task domains. These dialogs are given rich linguistic annotations by expert linguists for several types of reference mentions and named entity mentions, either of which can span multiple words, as well as for coreference links between mentions. The dialogs sometimes cross and blend domains, and the users exhibit complex task switching behavior such as re-initiating a previous task in the dialog by referencing the entities within it. The dataset contains a total of 8,429 dialogs with an average of 5.36 turns per dialog. We are releasing this dataset to encourage research in the field of coreference resolution, referring expression generation and identification within realistic, deep dialogs involving multiple domains. To demonstrate its utility, we also propose two baseline models for the downstream tasks: coreference resolution and referring expression generation.
Deep neural network models have played a critical role in sentiment analysis with promising results in the recent decade. One of the essential challenges, however, is how external sentiment knowledge can be effectively utilized. In this work, we propose a novel affection-driven approach to incorporating affective knowledge into neural network models. The affective knowledge is obtained in the form of a lexicon under the Affect Control Theory (ACT), which is represented by vectors of three-dimensional attributes in Evaluation, Potency, and Activity (EPA). The EPA vectors are mapped to an affective influence value and then integrated into Long Short-term Memory (LSTM) models to highlight affective terms. Experimental results show a consistent improvement of our approach over conventional LSTM models by 1.0% to 1.5% in accuracy on three large benchmark datasets. Evaluations across a variety of algorithms have also proven the effectiveness of leveraging affective terms for deep model enhancement.
The Alice Datasets are a set of datasets based on magnetic resonance data and electrophysiological data, collected while participants heard a story in English. Along with the datasets and the text of the story, we provide a variety of different linguistic and computational measures ranging from prosodic predictors to predictors capturing hierarchical syntactic information. These ecologically valid datasets can be easily reused to replicate prior work and to test new hypotheses about natural language comprehension in the brain.
Text-processing algorithms that annotate main components of a story-line are presently in great need of corpora and well-agreed annotation schemes. The Text World Theory of cognitive linguistics offers a model that generalizes a narrative structure in the form of world building elements (characters, time and space) as well as text worlds themselves and switches between them. We have conducted a survey on how text worlds and their elements are annotated in different projects and proposed our own annotation scheme and instructions. We tested them, first, on the science fiction story “We Can Remember It for You Wholesale” by Philip K. Dick. Then we corrected the guidelines and added computer annotation of verb forms with the purpose to get a higher raters’ agreement and tested them again on the short story “The Gift of the Magi” by O. Henry. As a result, the agreement among the three raters has risen. With due revision and tests, our annotation scheme and guidelines can be used for annotating narratives in corpora of literary texts, criminal evidence, teaching materials, quests, etc.
The paper relates to following ‘AC-hypotheses’: The articulatory code (AC) is a neural code exchanging multi-item messages between the short-term memory and cortical areas as the vSMC and STG. In these areas already neurons active in the presence of articulatory features have been measured. The AC codes the content of speech segmented in chunks and is the same for both modalities - speech perception and speech production. Each AC-message is related to a syllable. The items of each message relate to coordinated articulatory gestures composing the syllable. The mechanism to transport the AC and to segment the auditory signal is based on Ɵ/γ-oscillations, where a Ɵ-cycle has the duration of a Ɵ-syllable. The paper describes the findings from neuroscience, phonetics and the science of evolution leading to the AC-hypotheses. The paper proposes to verify the AC-hypotheses by measuring the activity of all ensembles of neurons coding and decoding the AC. Due to state of the art, the cortical measurements to be prepared, done and further processed need a high effort from scientists active in different areas. We propose to launch a project to produce cortical speech databases with cortical recordings synchronized with the speech signal allowing to decipher the articulatory code.
We recorded and preprocessed ZuCo 2.0, a new dataset of simultaneous eye-tracking and electroencephalography during natural reading and during annotation. This corpus contains gaze and brain activity data of 739 English sentences, 349 in a normal reading paradigm and 390 in a task-specific paradigm, in which the 18 participants actively search for a semantic relation type in the given sentences as a linguistic annotation task. This new dataset complements ZuCo 1.0 by providing experiments designed to analyze the differences in cognitive processing between natural reading and annotation. The data is freely available here: https://osf.io/2urht/.
Data from neuroscience and psychology suggest that sensorimotor cognition may be of central importance to language. Specifically, the linguistic structure of utterances referring to concrete actions may reflect the structure of the sensorimotor processing underlying the same action. To investigate this, we present the Linguistic, Kinematic and Gaze information in task descriptions Corpus (LKG-Corpus), comprising multimodal data on 13 humans, conducting take, put, and push actions, and describing these actions with 350 utterances. Recorded are audio, video, motion and eye-tracking data while participants perform an action and describe what they do. The dataset is annotated with orthographic transcriptions of utterances and information on: (a) gaze behaviours, (b) when a participant touched an object, (c) when an object was moved, (d) when a participant looked at the location s/he would next move the object to, (e) when the participant’s gaze was stable on an area. With the exception of the annotation of stable gaze, all annotations were performed manually. With the LKG-Corpus, we present a dataset that integrates linguistic, kinematic and gaze data with an explicit focus on relations between action and language. On this basis, we outline applications of the dataset to both basic and applied research.
We present the ACQDIV corpus database and aggregation pipeline, a tool developed as part of the European Research Council (ERC) funded project ACQDIV, which aims to identify the universal cognitive processes that allow children to acquire any language. The corpus database represents 15 corpora from 14 typologically maximally diverse languages. Here we give an overview of the project, database, and our extensible software package for adding more corpora to the current language sample. Lastly, we discuss how we use the corpus database to mine for universal patterns in child language acquisition corpora and we describe avenues for future research.
This article presents a resource that links WordNet, the widely known lexical and semantic database, and Arasaac, the largest freely available database of pictograms. Pictograms are a tool that is more and more used by people with cognitive or communication disabilities. However, they are mainly used manually via workbooks, whereas caregivers and families would like to use more automated tools (use speech to generate pictograms, for example). In order to make it possible to use pictograms automatically in NLP applications, we propose a database that links them to semantic knowledge. This resource is particularly interesting for the creation of applications that help people with cognitive disabilities, such as text-to-picto, speech-to-picto, picto-to-speech... In this article, we explain the needs for this database and the problems that have been identified. Currently, this resource combines approximately 800 pictograms with their corresponding WordNet synsets and it is accessible both through a digital collection and via an SQL database. Finally, we propose a method with associated tools to make our resource language-independent: this method was applied to create a first text-to-picto prototype for the French language. Our resource is distributed freely under a Creative Commons license at the following URL: https://github.com/getalp/Arasaac-WN.
We consider the orthographic neighborhood effect: the effect that words with more orthographic similarity to other words are read faster. The neighborhood effect serves as an important control variable in psycholinguistic studies of word reading, and explains variance in addition to word length and word frequency. Following previous work, we model the neighborhood effect as the average distance to neighbors in feature space for three feature sets: slots, character ngrams and skipgrams. We optimize each of these feature sets and find evidence for language-independent optima, across five megastudy corpora from five alphabetic languages. Additionally, we show that weighting features using the inverse of mutual information (MI) improves the neighborhood effect significantly for all languages. We analyze the inverse feature weighting, and show that, across languages, grammatical morphemes get the lowest weights. Finally, we perform the same experiments on Korean Hangul, a non-alphabetic writing system, where we find the opposite results: slower responses as a function of denser neighborhoods, and a negative effect of inverse feature weighting. This raises the question of whether this is a cognitive effect, or an effect of the way we represent Hangul orthography, and indicates more research is needed.
The purpose of this paper is twofold:  to introduce, to our knowledge, the largest available resource of keystroke logging (KSL) data generated by Etherpad (https://etherpad.org/), an open-source, web-based collaborative real-time editor, that captures the dynamics of second language (L2) production and  to relate the behavioral data from KSL to indices of syntactic and lexical complexity of the texts produced obtained from a tool that implements a sliding window approach capturing the progression of complexity within a text. We present the procedures and measures developed to analyze a sample of 14,913,009 keystrokes in 3,454 texts produced by 512 university students (upper-intermediate to advanced L2 learners of English) (95,354 sentences and 18,32,027 words) aiming to achieve a better alignment between keystroke-logging measures and underlying cognitive processes, on the one hand, and L2 writing performance measures, on the other hand. The resource introduced in this paper is a reflection of increasing recognition of the urgent need to obtain ecologically valid data that have the potential to transform our current understanding of mechanisms underlying the development of literacy (reading and writing) skills.
The past decade has witnessed the happy marriage between natural language processing (NLP) and the cognitive science of language. Moreover, given the historical relationship between biological and artificial neural networks, the advent of deep learning has re-sparked strong interests in the fusion of NLP and the neuroscience of language. Importantly, this inter-fertilization between NLP, on one hand, and the cognitive (neuro)science of language, on the other, has been driven by the language resources annotated with human language processing data. However, there remain several limitations with those language resources on annotations, genres, languages, etc. In this paper, we describe the design of a novel language resource called BCCWJ-EEG, the Balanced Corpus of Contemporary Written Japanese (BCCWJ) experimentally annotated with human electroencephalography (EEG). Specifically, after extensively reviewing the language resources currently available in the literature with special focus on eye-tracking and EEG, we summarize the details concerning (i) participants, (ii) stimuli, (iii) procedure, (iv) data preprocessing, (v) corpus evaluation, (vi) resource release, and (vii) compilation schedule. In addition, potential applications of BCCWJ-EEG to neuroscience and NLP will also be discussed.
In modern linguistics and psycholinguistics speech disfluencies in real fluent speech are a well-known phenomenon. But it’s not still clear which components of brain systems are involved into its comprehension in a listener’s brain. In this paper we provide a pilot neuroimaging study of the possible neural correlates of speech disfluencies perception, using a combination of the corpus and functional magnetic-resonance imaging (fMRI) methods. Special technical procedure of selecting stimulus material from Russian multichannel corpus RUPEX allowed to create fragments in terms of requirements for the fMRI BOLD temporal resolution. They contain isolated speech disfluencies and their clusters. Also, we used the referential task for participants fMRI scanning. As a result, it was demonstrated that annotated multichannel corpora like RUPEX can be an important resource for experimental research in interdisciplinary fields. Thus, different aspects of communication can be explored through the prism of brain activation.
The NAIST Lang-8 Learner Corpora (Lang-8 corpus) is one of the largest second-language learner corpora. The Lang-8 corpus is suitable as a training dataset for machine translation-based grammatical error correction systems. However, it is not suitable as an evaluation dataset because the corrected sentences sometimes include inappropriate sentences. Therefore, we created and released an evaluation corpus for correcting grammatical errors made by learners of Japanese as a Second Language (JSL). As our corpus has less noise and its annotation scheme reflects the characteristics of the dataset, it is ideal as an evaluation corpus for correcting grammatical errors in sentences written by JSL learners. In addition, we applied neural machine translation (NMT) and statistical machine translation (SMT) techniques to correct the grammar of the JSL learners’ sentences and evaluated their results using our corpus. We also compared the performance of the NMT system with that of the SMT system.
Information extraction from unstructured texts plays a vital role in the field of natural language processing. Although there has been extensive research into each information extraction task (i.e., entity linking, coreference resolution, and relation extraction), data are not available for a continuous and coherent evaluation of all information extraction tasks in a comprehensive framework. Given that each task is performed and evaluated with a different dataset, analyzing the effect of the previous task on the next task with a single dataset throughout the information extraction process is impossible. This paper aims to propose a Korean information extraction initiative point and promote research in this field by presenting crowdsourcing data collected for four information extraction tasks from the same corpus and the training and evaluation results for each task of a state-of-the-art model. These machine learning data for Korean information extraction are the first of their kind, and there are plans to continuously increase the data volume. The test results will serve as an initiative result for each Korean information extraction task and are expected to serve as a comparison target for various studies on Korean information extraction using the data collected in this study.
Resolving Indirect Speech Acts (ISAs), in which the intended meaning of an utterance is not identical to its literal meaning, is essential to enabling the participation of intelligent systems in peoples’ everyday lives. Especially challenging are those cases in which the interpretation of such ISAs depends on context. To test a system’s ability to perform ISA resolution we need a corpus, but developing such a corpus is difficult, especialy given the contex-dependent requirement. This paper addresses the difficult problems of constructing a corpus of ISAs, taking inspiration from relevant work in using corpora for reasoning tasks. We present a formal representation of ISA Schemas required for such testing, including a measure of the difficulty of a particular schema. We develop an approach to authoring these schemas using corpus analysis and crowdsourcing, to maximize realism and minimize the amount of expert authoring needed. Finally, we describe several characteristics of collected data, and potential future work.
The quality estimation of artifacts generated by creators via crowdsourcing has great significance for the construction of a large-scale data resource. A common approach to this problem is to ask multiple reviewers to evaluate the same artifacts. However, the commonly used majority voting method to aggregate reviewers’ evaluations does not work effectively for partially subjective or purely subjective tasks because reviewers’ sensitivity and bias of evaluation tend to have a wide variety. To overcome this difficulty, we propose a probabilistic model for subjective classification tasks that incorporates the qualities of artifacts as well as the abilities and biases of creators and reviewers as latent variables to be jointly inferred. We applied this method to the partially subjective task of speech classification into the following four attitudes: agreement, disagreement, stalling, and question. The result shows that the proposed method estimates the quality of speech more effectively than a vote aggregation, measured by correlation with a fine-grained classification by experts.
Using current methods, the construction of multilingual resources in FrameNet is an expensive and complex task. While crowdsourcing is a viable alternative, it is difficult to include non-native English speakers in such efforts as they often have difficulty with English-based FrameNet tools. In this work, we investigated cross-lingual issues in crowdsourcing approaches for multilingual FrameNets, specifically in the context of the newly constructed Korean FrameNet. To accomplish this, we evaluated the effectiveness of various crowdsourcing settings whereby certain types of information are provided to workers, such as English definitions in FrameNet or translated definitions. We then evaluated whether the crowdsourced results accurately captured the meaning of frames both cross-culturally and cross-linguistically, and found that by allowing the crowd workers to make intuitive choices, they achieved a quality comparable to that of trained FrameNet experts (F1 > 0.75). The outcomes of this work are now publicly available as a new release of Korean FrameNet 1.1.
The intrinsic and extrinsic quality evaluation is an essential part of the summary evaluation methodology usually conducted in a traditional controlled laboratory environment. However, processing large text corpora using these methods reveals expensive from both the organizational and the financial perspective. For the first time, and as a fast, scalable, and cost-effective alternative, we propose micro-task crowdsourcing to evaluate both the intrinsic and extrinsic quality of query-based extractive text summaries. To investigate the appropriateness of crowdsourcing for this task, we conduct intensive comparative crowdsourcing and laboratory experiments, evaluating nine extrinsic and intrinsic quality measures on 5-point MOS scales. Correlating results of crowd and laboratory ratings reveals high applicability of crowdsourcing for the factors overall quality, grammaticality, non-redundancy, referential clarity, focus, structure & coherence, summary usefulness, and summary informativeness. Further, we investigate the effect of the number of repetitions of assessments on the robustness of mean opinion score of crowd ratings, measured against the increase of correlation coefficients between crowd and laboratory. Our results suggest that the optimal number of repetitions in crowdsourcing setups, in which any additional repetitions do no longer cause an adequate increase of overall correlation coefficients, lies between seven and nine for intrinsic and extrinsic quality factors.
Temples are an integral part of culture and heritage of India and are centers of religious practice for practicing Hindus. A scientific study of temples can reveal valuable insights into Indian culture and heritage. However to the best of our knowledge, learning resources that aid such a study are either not publicly available or non-existent. In this endeavour we present our initial efforts to create a corpus of Hindu temples in India. In this paper, we present a simple, re-usable platform that creates temple corpus from web text on temples. Curation is improved using classifiers trained on textual data in Wikipedia articles on Hindu temples. The training data is verified by human volunteers. The temple corpus consists of 4933 high accuracy facts about 573 temples. We make the corpus and the platform freely available. We also test the re-usability of the platform by creating a corpus of museums in India. We believe the temple corpus will aid scientific study of temples and the platform will aid in construction of similar corpuses. We believe both these will significantly contribute in promoting research on culture and heritage of a region.
This work collects and studies Chinese readers’ veridicality judgments to news events (whether an event is viewed as happening or not). For instance, in “The FBI alleged in court documents that Zazi had admitted having a handwritten recipe for explosives on his computer”, do people believe that Zazi had a handwritten recipe for explosives? The goal is to observe the pragmatic behaviors of linguistic features under context which affects readers in making veridicality judgments. Exploring from the datasets, it is found that features such as event-selecting predicates (ESP), modality markers, adverbs, temporal information, and statistics have an impact on readers’ veridicality judgments. We further investigated that modality markers with high certainty do not necessarily trigger readers to have high confidence in believing an event happened. Additionally, the source of information introduced by an ESP presents low effects to veridicality judgments, even when an event is attributed to an authority (e.g. “The FBI”). A corpus annotated with Chinese readers’ veridicality judgments is released as the Chinese PragBank for further analysis.
We introduce in this paper a generic approach to combine implicit crowdsourcing and language learning in order to mass-produce language resources (LRs) for any language for which a crowd of language learners can be involved. We present the approach by explaining its core paradigm that consists in pairing specific types of LRs with specific exercises, by detailing both its strengths and challenges, and by discussing how much these challenges have been addressed at present. Accordingly, we also report on on-going proof-of-concept efforts aiming at developing the first prototypical implementation of the approach in order to correct and extend an LR called ConceptNet based on the input crowdsourced from language learners. We then present an international network called the European Network for Combining Language Learning with Crowdsourcing Techniques (enetCollect) that provides the context to accelerate the implementation of this generic approach. Finally, we exemplify how it can be used in several language learning scenarios to produce a multitude of NLP resources and how it can therefore alleviate the long-standing NLP issue of the lack of LRs.
Given the limited size of existing idiom corpora, we aim to enable progress in automatic idiom processing and linguistic analysis by creating the largest-to-date corpus of idioms for English. Using a fixed idiom list, automatic pre-extraction, and a strictly controlled crowdsourced annotation procedure, we show that it is feasible to build a high-quality corpus comprising more than 50K instances, an order of a magnitude larger than previous resources. Crucial ingredients of crowdsourcing were the selection of crowdworkers, clear and comprehensive instructions, and an interface that breaks down the task in small, manageable steps. Analysis of the resulting corpus revealed strong effects of genre on idiom distribution, providing new evidence for existing theories on what influences idiom usage. The corpus also contains rich metadata, and is made publicly available.
Large corpora of task-based and open-domain conversational dialogues are hugely valuable in the field of data-driven dialogue systems. Crowdsourcing platforms, such as Amazon Mechanical Turk, have been an effective method for collecting such large amounts of data. However, difficulties arise when task-based dialogues require expert domain knowledge or rapid access to domain-relevant information, such as databases for tourism. This will become even more prevalent as dialogue systems become increasingly ambitious, expanding into tasks with high levels of complexity that require collaboration and forward planning, such as in our domain of emergency response. In this paper, we propose CRWIZ: a framework for collecting real-time Wizard of Oz dialogues through crowdsourcing for collaborative, complex tasks. This framework uses semi-guided dialogue to avoid interactions that breach procedures and processes only known to experts, while enabling the capture of a wide variety of interactions.
Named Entity Recognition (NER) is an essential component of many Natural Language Processing pipelines. However, building these language dependent models requires large amounts of annotated data. Crowdsourcing emerged as a scalable solution to collect and enrich data in a more time-efficient manner. To manage these annotations at scale, it is important to predict completion timelines and compute fair pricing for workers in advance. To achieve these goals, we need to know how much effort will be taken to complete each task. In this paper, we investigate which variables influence the time spent on a named entity annotation task by a human. Our results are two-fold: first, the understanding of the effort-impacting factors which we divided into cognitive load and input length; and second, the performance of the prediction itself. On the latter, through model adaptation and feature engineering, we attained a Root Mean Squared Error (RMSE) of 25.68 words per minute with a Nearest Neighbors model.
In this work, we report on a crowdsourcing experiment conducted using the V-TREL vocabulary trainer which is accessed via a Telegram chatbot interface to gather knowledge on word relations suitable for expanding ConceptNet. V-TREL is built on top of a generic architecture implementing the implicit crowdsourding paradigm in order to offer vocabulary training exercises generated from the commonsense knowledge-base ConceptNet and – in the background – to collect and evaluate the learners’ answers to extend ConceptNet with new words. In the experiment about 90 university students learning English at C1 level, based on Common European Framework of Reference for Languages (CEFR), trained their vocabulary with V-TREL over a period of 16 calendar days. The experiment allowed to gather more than 12,000 answers from learners on different question types. In this paper we present in detail the experimental setup and the outcome of the experiment, which indicates the potential of our approach for both crowdsourcing data as well as fostering vocabulary skills.
The objective of this research is to estimate multidimensional subjective ratings of the reading performance of young readers from signal-based objective measures. We here combine linguistic features (number of correct words, repetitions, deletions, insertions uttered per minute . . . ) with phonetic features. Expressivity is particularly difficult to predict since there is no unique golden standard. We here propose a novel framework for performing such an estimation that exploits multiple references performed by adults and demonstrate its efficiency using recordings of 273 pupils.
LARA (Learning and Reading Assistant) is an open source platform whose purpose is to support easy conversion of plain texts into multimodal online versions suitable for use by language learners. This involves semi-automatically tagging the text, adding other annotations and recording audio. The platform is suitable for creating texts in multiple languages via crowdsourcing techniques that can be used for teaching a language via reading and listening. We present results of initial experiments by various collaborators where we measure the time required to produce substantial LARA resources, up to the length of short novels, in Dutch, English, Farsi, French, German, Icelandic, Irish, Swedish and Turkish. The first results are encouraging. Although there are some startup problems, the conversion task seems manageable for the languages tested so far. The resulting enriched texts are posted online and are freely available in both source and compiled form.
We present an annotation scheme and a dataset of teacher feedback provided for texts written by non-native speakers of English. The dataset consists of student-written sentences in their original and revised versions with teacher feedback provided for the errors. Feedback appears both in the form of open-ended comments and error category tags. We focus on a specific error type, namely linking adverbial (e.g. however, moreover) errors. The dataset has been annotated for two aspects: (i) revision outcome establishing whether the re-written student sentence was correct and (ii) directness, indicating whether teachers provided explicitly the correction in their feedback. This dataset allows for studies around the characteristics of teacher feedback and how these influence students’ revision outcome. We describe the data preparation process and we present initial statistical investigations regarding the effect of different feedback characteristics on revision outcome. These show that open-ended comments and mitigating expressions appear in a higher proportion of successful revisions than unsuccessful ones, while directness and metalinguistic terms have no effect. Given that the use of this type of data is relatively unexplored in natural language processing (NLP) applications, we also report some observations and challenges when working with feedback data.
In this paper, we report on datasets that we created for research in feedback comment generation — a task of automatically generating feedback comments such as a hint or an explanatory note for writing learning. There has been almost no such corpus open to the public and accordingly there has been a very limited amount of work on this task. In this paper, we first discuss the principle and guidelines for feedback comment annotation. Then, we describe two corpora that we have manually annotated with feedback comments (approximately 50,000 general comments and 6,700 on preposition use). A part of the annotation results is now available on the web, which will facilitate research in feedback comment generation
The Common European Framework of Reference for Languages (CEFR) defines six levels of learner proficiency, and links them to particular communicative abilities. The CEFRLex project aims at compiling lexical resources that link single words and multi-word expressions to particular CEFR levels. The resources are thought to reflect second language learner needs as they are compiled from CEFR-graded textbooks and other learner-directed texts. In this work, we investigate the applicability of CEFRLex resources for building language learning applications. Our main concerns were that vocabulary in language learning materials might be sparse, i.e. that not all vocabulary items that belong to a particular level would also occur in materials for that level, and, on the other hand, that vocabulary items might be used on lower-level materials if required by the topic (e.g. with a simpler paraphrasing or translation). Our results indicate that the English CEFRLex resource is in accordance with external resources that we jointly employ as gold standard. Together with other values obtained from monolingual and parallel corpora, we can indicate which entries need to be adjusted to obtain values that are even more in line with this gold standard. We expect that this finding also holds for the other languages
The use of Augmented Reality (AR) in teaching and learning contexts for language is still young. The ideas are endless, the concrete educational offers available emerge only gradually. Educational opportunities that were unthinkable a few years ago are now feasible. We present a concrete realization: an executable application for mobile devices with which users can explore their environment interactively in different languages. The software recognizes up to 1000 objects in the user’s environment using a deep learning method based on Convolutional Neural Networks and names this objects accordingly. Using Augmented Reality the objects are superimposed with 3D information in different languages. By switching the languages, the user is able to interactively discover his surrounding everyday items in all languages. The application is available as Open Source.
Revision plays a major role in writing and the analysis of writing processes. Revisions can be analyzed using a product-oriented approach (focusing on a finished product, the text that has been produced) or a process-oriented approach (focusing on the process that the writer followed to generate this product). Although several language resources exist for the product-oriented approach to revisions, there are hardly any resources available yet for an in-depth analysis of the process of revisions. Therefore, we provide an extensive dataset on revisions made during writing (accessible via https://hdl.handle.net/10411/VBDYGX). This dataset is based on keystroke data and eye tracking data of 65 students from a variety of backgrounds (undergraduate and graduate English as a first language and English as a second language students) and a variety of tasks (argumentative text and academic abstract). In total, 7,120 revisions were identified in the dataset. For each revision, 18 features have been manually annotated and 31 features have been automatically extracted. As a case study, we show two potential use cases of the dataset. In addition, future uses of the dataset are described.
This paper introduces a new web system that integrates English Grammatical Error Detection (GED) and course-specific stylistic guidelines to automatically review and provide feedback on student assignments. The system is being developed as a pedagogical tool for English Scientific Writing. It uses both general NLP methods and high precision parsers to check student assignments before they are submitted for grading. Instead of generalized error detection, our system aims to identify, with high precision, specific classes of problems that are known to be common among engineering students. Rather than correct the errors, our system generates constructive feedback to help students identify and correct them on their own. A preliminary evaluation of the system’s in-class performance has shown measurable improvements in the quality of student assignments.
This paper describes “TLT-school” a corpus of speech utterances collected in schools of northern Italy for assessing the performance of students learning both English and German. The corpus was recorded in the years 2017 and 2018 from students aged between nine and sixteen years, attending primary, middle and high school. All utterances have been scored, in terms of some predefined proficiency indicators, by human experts. In addition, most of utterances recorded in 2017 have been manually transcribed carefully. Guidelines and procedures used for manual transcriptions of utterances will be described in detail, as well as results achieved by means of an automatic speech recognition system developed by us. Part of the corpus is going to be freely distributed to scientific community particularly interested both in non-native speech recognition and automatic assessment of second language proficiency.
We present the first version of the longitudinal Revita Learner Corpus (ReLCo), for Russian. In contrast to traditional learner corpora, ReLCo is collected and annotated fully automatically, while students perform exercises using the Revita language-learning platform. The corpus currently contains 8 422 sentences exhibiting several types of errors—grammatical, lexical, orthographic, etc.—which were committed by learners during practice and were automatically annotated by Revita. The corpus provides valuable information about patterns of learner errors and can be used as a language resource for a number of research tasks, while its creation is much cheaper and faster than for traditional learner corpora. A crucial advantage of ReLCo that it grows continually while learners practice with Revita, which opens the possibility of creating an unlimited learner resource with longitudinal data collected over time. We make the pilot version of the Russian ReLCo publicly available.
The paper presents quality focused approach to a learner corpus development. The methodology was developed with multiple design considerations put in place to make the annotation process easier and at the same time reduce the amount of mistakes that could be introduced due to inconsistent text correction or carelessness. The approach suggested in this paper consists of multiple parts: comparison of digitized texts by several annotators, text correction, automated morphological analysis, and manual review of annotations. The described approach is used to create Latvian Language Learner corpus (LaVA) which is part of a currently ongoing project Development of Learner corpus of Latvian: methods, tools and applications.
Automated writing evaluation is a popular research field, but the main focus has been on evaluating argumentative essays. In this paper, we consider a different genre, namely précis texts. A précis is a written text that provides a coherent summary of main points of a spoken or written text. We present a corpus of English précis texts which all received a grade assigned by a highly-experienced English language teacher and were subsequently annotated following an exhaustive error typology. With this corpus we trained a machine learning model which relies on a number of linguistic, automatic summarization and AWE features. Our results reveal that this model is able to predict the grade of précis texts with only a moderate error margin.
Natural Language Image Editing (NLIE) aims to use natural language instructions to edit images. Since novices are inexperienced with image editing techniques, their instructions are often ambiguous and contain high-level abstractions which require complex editing steps. Motivated by this inexperience aspect, we aim to smooth the learning curve by teaching the novices to edit images using low-level command terminologies. Towards this end, we develop a task-oriented dialogue system to investigate low-level instructions for NLIE. Our system grounds language on the level of edit operations, and suggests options for users to choose from. Though compelled to express in low-level terms, user evaluation shows that 25% of users found our system easy-to-use, resonating with our motivation. Analysis shows that users generally adapt to utilizing the proposed low-level language interface. We also identified object segmentation as the key factor to user satisfaction. Our work demonstrates advantages of low-level, direct language-action mapping approach that can be applied to other problem domains beyond image editing such as audio editing or industrial design.
For every patient’s visit to a clinician, a clinical note is generated documenting their medical conversation, including complaints discussed, treatments, and medical plans. Despite advances in natural language processing, automating clinical note generation from a clinic visit conversation is a largely unexplored area of research. Due to the idiosyncrasies of the task, traditional methods of corpus creation are not effective enough approaches for this problem. In this paper, we present an annotation methodology that is content- and technique- agnostic while associating note sentences to sets of dialogue sentences. The sets can further be grouped with higher order tags to mark sets with related information. This direct linkage from input to output decouples the annotation from specific language understanding or generation strategies. Here we provide data statistics and qualitative analysis describing the unique annotation challenges. Given enough annotated data, such a resource would support multiple modeling methods including information extraction with template language generation, information retrieval type language generation, or sequence to sequence modeling.
MultiWOZ 2.0 (Budzianowski et al., 2018) is a recently released multi-domain dialogue dataset spanning 7 distinct domains and containing over 10,000 dialogues. Though immensely useful and one of the largest resources of its kind to-date, MultiWOZ 2.0 has a few shortcomings. Firstly, there are substantial noise in the dialogue state annotations and dialogue utterances which negatively impact the performance of state-tracking models. Secondly, follow-up work (Lee et al., 2019) has augmented the original dataset with user dialogue acts. This leads to multiple co-existent versions of the same dataset with minor modifications. In this work we tackle the aforementioned issues by introducing MultiWOZ 2.1. To fix the noisy state annotations, we use crowdsourced workers to re-annotate state and utterances based on the original utterances in the dataset. This correction process results in changes to over 32% of state annotations across 40% of the dialogue turns. In addition, we fix 146 dialogue utterances by canonicalizing slot values in the utterances to the values in the dataset ontology. To address the second problem, we combined the contributions of the follow-up works into MultiWOZ 2.1. Hence, our dataset also includes user dialogue acts as well as multiple slot descriptions per dialogue state slot. We then benchmark a number of state-of-the-art dialogue state tracking models on the MultiWOZ 2.1 dataset and show the joint state tracking performance on the corrected state annotations. We are publicly releasing MultiWOZ 2.1 to the community, hoping that this dataset resource will allow for more effective models across various dialogue subproblems to be built in the future.
Recommendation systems aim at facilitating information retrieval for users by taking into account their preferences. Based on previous user behaviour, such a system suggests items or provides information that a user might like or find useful. Nonetheless, how to provide suggestions is still an open question. Depending on the way a recommendation is communicated influences the user’s perception of the system. This paper presents an empirical study on the effects of proactive dialogue strategies on user acceptance. Therefore, an explicit strategy based on user preferences provided directly by the user, and an implicit proactive strategy, using autonomously gathered information, are compared. The results show that proactive dialogue systems significantly affect the perception of human-computer interaction. Although no significant differences are found between implicit and explicit strategies, proactivity significantly influences the user experience compared to reactive system behaviour. The study contributes new insights to the human-agent interaction and the voice user interface design. Furthermore, we discover interesting tendencies that motivate futurework.
Conversational Question Answering (CQA) systems meet user information needs by having conversations with them, where answers to the questions are retrieved from text. There exist a variety of datasets for English, with tens of thousands of training examples, and pre-trained language models have allowed to obtain impressive results. The goal of our research is to test the performance of CQA systems under low-resource conditions which are common for most non-English languages: small amounts of native annotations and other limitations linked to low resource languages, like lack of crowdworkers or smaller wikipedias. We focus on the Basque language, and present the first non-English CQA dataset and results. Our experiments show that it is possible to obtain good results with low amounts of native data thanks to cross-lingual transfer, with quality comparable to those obtained for English. We also discovered that dialogue history models are not directly transferable to another language, calling for further research. The dataset is publicly available.
There are high expectations for multimodal dialog systems that can make natural small talk with facial expressions, gestures, and gaze actions as next-generation dialog-based systems. Two important roles of the chat-talk system are keeping the user engaged and establishing rapport. Many studies have conducted user evaluations of such systems, some of which reported that considering the relationship with the user is an effective way to improve the subjective evaluation. To facilitate research of such dialog systems, we are currently constructing a large-scale multimodal dialog corpus focusing on the relationship between speakers. In this paper, we describe the data collection and annotation process, and analysis of the corpus collected in the early stage of the project. This corpus contains 19,303 utterances (10 hours) from 19 pairs of participants. A dialog act tag is annotated to each utterance by two annotators. We compare the frequency and the transition probability of the tags between different closeness levels to help construct a dialog system for establishing a relationship with the user.
An important objective in health-technology is the ability to gather information about people’s well-being. Structured interviews can be used to obtain this information, but are time-consuming and not scalable. Questionnaires provide an alternative way to extract such information, though typically lack depth. In this paper, we present our first prototype of the BLISS agent, an artificial intelligent agent which intends to automatically discover what makes people happy and healthy. The goal of Behaviour-based Language-Interactive Speaking Systems (BLISS) is to understand the motivations behind people’s happiness by conducting a personalized spoken dialogue based on a happiness model. We built our first prototype of the model to collect 55 spoken dialogues, in which the BLISS agent asked questions to users about their happiness and well-being. Apart from a description of the BLISS architecture, we also provide details about our dataset, which contains over 120 activities and 100 motivations and is made available for usage.
Human conversations are complicated and building a human-like dialogue agent is an extremely challenging task. With the rapid development of deep learning techniques, data-driven models become more and more prevalent which need a huge amount of real conversation data. In this paper, we construct a large-scale real scenario Chinese E-commerce conversation corpus, JDDC, with more than 1 million multi-turn dialogues, 20 million utterances, and 150 million words. The dataset reflects several characteristics of human-human conversations, e.g., goal-driven, and long-term dependency among the context. It also covers various dialogue types including task-oriented, chitchat and question-answering. Extra intent information and three well-annotated challenge sets are also provided. Then, we evaluate several retrieval-based and generative models to provide basic benchmark performance on the JDDC corpus. And we hope JDDC can serve as an effective testbed and benefit the development of fundamental research in dialogue task.
Cheese! is a conversational corpus. It consists of 11 French face-to-face conversations lasting around 15 minutes each. Cheese! is a duplication of an American corpus (ref) in order to conduct a cross-cultural comparison of participants’ smiling behavior in humorous and non-humorous sequences in American English and French conversations. In this article, the methodology used to collect and enrich the corpus is presented: experimental protocol, technical choices, transcription, semi-automatic annotations, manual annotations of smiling and humor. An exploratory study investigating the links between smile and humor is then proposed. Based on the analysis of two interactions, two questions are asked: (1) Does smile frame humor? (2) Does smile has an impact on its success or failure? If the experimental design of Cheese! has been elaborated to study specifically smiles and humor in conversations, the high quality of the dataset obtained, and the methodology used are also replicable and can be applied to analyze many other conversational activities and other multimodal modalities.
Time-Offset Interaction Applications (TOIAs) are systems that simulate face-to-face conversations between humans and digital human avatars recorded in the past. Developing a well-functioning TOIA involves several research areas: artificial intelligence, human-computer interaction, natural language processing, question answering, and dialogue systems. The first challenges are to define a sensible methodology for data collection and to create useful data sets for training the system to retrieve the best answer to a user’s question. In this paper, we present three main contributions: a methodology for creating the knowledge base for a TOIA, a dialogue corpus, and baselines for single-turn answer retrieval. We develop the methodology using a two-step strategy. First, we let the avatar maker list pairs by intuition, guessing what possible questions a user may ask to the avatar. Second, we record actual dialogues between random individuals and the avatar-maker. We make the Margarita Dialogue Corpus available to the research community. This corpus comprises the knowledge base in text format, the video clips for each answer, and the annotated dialogues.
Nowadays Personal Assistants (PAs) are available in multiple environments and become increasingly popular to use via voice. Therefore, we aim to provide proactive PA suggestions to car drivers via speech. These suggestions should be neither obtrusive nor increase the drivers’ cognitive load, while enhancing user experience. To assess these factors, we conducted a usability study in which 42 participants perceive proactive voice output in a Wizard-of-Oz study in a driving simulator. Traffic density was varied during a highway drive and it included six in-car-specific use cases. The latter were presented by a proactive voice assistant and in a non-proactive control condition. We assessed the users’ subjective cognitive load and their satisfaction in different questionnaires during the interaction with both PA variants. Furthermore, we analyze the user reactions: both regarding their content and the elapsed response times to PA actions. The results show that proactive assistant behavior is rated similarly positive as non-proactive behavior. Furthermore, the participants agreed to 73.8% of proactive suggestions. In line with previous research, driving-relevant use cases receive the best ratings, here we reach 82.5% acceptance. Finally, the users reacted significantly faster to proactive PA actions, which we interpret as less cognitive load compared to non-proactive behavior.
Expressing emotion is known as an efficient way to persuade one’s dialogue partner to accept one’s claim or proposal. Emotional expression in speech can express the speaker’s emotion more directly than using only emotion expression in the text, which will lead to a more persuasive dialogue. In this paper, we built a speech dialogue corpus in a persuasive scenario that uses emotional expressions to build a persuasive dialogue system with emotional expressions. We extended an existing text dialogue corpus by adding variations of emotional responses to cover different combinations of broad dialogue context and a variety of emotional states by crowd-sourcing. Then, we recorded emotional speech consisting of of collected emotional expressions spoken by a voice actor. The experimental results indicate that the collected emotional expressions with their speeches have higher emotional expressiveness for expressing the system’s emotion to users.
Group cohesion is an emergent phenomenon that describes the tendency of the group members’ shared commitment to group tasks and the interpersonal attraction among them. This paper presents a multimodal analysis of group cohesion using a corpus of multi-party interactions. We utilize 16 two-minute segments annotated with cohesion from the AMI corpus. We define three layers of modalities: non-verbal social cues, dialogue acts and interruptions. The initial analysis is performed at the individual level and later, we combine the different modalities to observe their impact on perceived level of cohesion. Results indicate that occurrence of laughter and interruption are higher in high cohesive segments. We also observe that, dialogue acts and head nods did not have an impact on the level of cohesion by itself. However, when combined there was an impact on the perceived level of cohesion. Overall, the analysis shows that multimodal cues are crucial for accurate analysis of group cohesion.
Dialogue systems for interaction with humans have been enjoying increased popularity in the research and industry fields. To this day, the best way to estimate their success is through means of human evaluation and not automated approaches, despite the abundance of work done in the field. In this paper, we investigate the effectiveness of perceiving dialogue evaluation as an anomaly detection task. The paper looks into four dialogue modeling approaches and how their objective functions correlate with human annotation scores. A high-level perspective exhibits negative results. However, a more in-depth look shows some potential for using anomaly detection for evaluating dialogues.
We present an approach to evaluate argument search techniques in view of their use in argumentative dialogue systems by assessing quality aspects of the retrieved arguments. To this end, we introduce a dialogue system that presents arguments by means of a virtual avatar and synthetic speech to users and allows them to rate the presented content in four different categories (Interesting, Convincing, Comprehensible, Relation). The approach is applied in a user study in order to compare two state of the art argument search engines to each other and with a system based on traditional web search. The results show a significant advantage of the two search engines over the baseline. Moreover, the two search engines show significant advantages over each other in different categories, thereby reflecting strengths and weaknesses of the different underlying techniques.
The recognition and automatic annotation of temporal expressions (e.g. “Add an event for tomorrow evening at eight to my calendar”) is a key module for AI voice assistants, in order to allow them to interact with apps (for example, a calendar app). However, in the NLP literature, research on temporal expressions has focused mostly on data from the news, from the clinical domain, and from social media. The voice assistant domain is very different than the typical domains that have been the focus of work on temporal expression identification, thus requiring a dedicated data collection. We present a crowdsourcing method for eliciting natural-language commands containing temporal expressions for an AI voice assistant, by using pictures and scenario descriptions. We annotated the elicited commands (480) as well as the commands in the Snips dataset following the TimeML/TIMEX3 annotation guidelines, reaching a total of 1188 annotated commands. The commands can be later used to train the NLU components of an AI voice assistant.
ISO 24617-2, the ISO standard for dialog act annotation, sets the ground for more comparable research in the area. However, the amount of data annotated according to it is still reduced, which impairs the development of approaches for automatic recognition. In this paper, we describe a mapping of the original dialog act labels of the LEGO corpus, which have been neglected, into the communicative functions of the standard. Although this does not lead to a complete annotation according to the standard, the 347 dialogs provide a relevant amount of data that can be used in the development of automatic communicative function recognition approaches, which may lead to a wider adoption of the standard. Using the 17 English dialogs of the DialogBank as gold standard, our preliminary experiments have shown that including the mapped dialogs during the training phase leads to improved performance while recognizing communicative functions in the Task dimension.
We present a neural network approach to estimate the communication style of spoken interaction, namely the stylistic variations elaborateness and directness, and investigate which type of input features to the estimator are necessary to achive good performance. First, we describe our annotated corpus of recordings in the health care domain and analyse the corpus statistics in terms of agreement, correlation and reliability of the ratings. We use this corpus to estimate the elaborateness and the directness of each utterance. We test different feature sets consisting of dialogue act features, grammatical features and linguistic features as input for our classifier and perform classification in two and three classes. Our classifiers use only features that can be automatically derived during an ongoing interaction in any spoken dialogue system without any prior annotation. Our results show that the elaborateness can be classified by only using the dialogue act and the amount of words contained in the corresponding utterance. The directness is a more difficult classification task and additional linguistic features in form of word embeddings improve the classification results. Afterwards, we run a comparison with a support vector machine and a recurrent neural network classifier.
ISO standard 24617-2 for dialogue act annotation, established in 2012, has in the past few years been used both in corpus annotation and in the design of components for spoken and multimodal dialogue systems. This has brought some inaccuracies and undesirbale limitations of the standard to light, which are addressed in a proposed second edition. This second edition allows a more accurate annotation of dependence relations and rhetorical relations in dialogue. Following the ISO 24617-4 principles of semantic annotation, and borrowing ideas from EmotionML, a triple-layered plug-in mechanism is introduced which allows dialogue act descriptions to be enriched with information about their semantic content, about accompanying emotions, and other information, and allows the annotation scheme to be customised by adding application-specific dialogue act types.
This paper describes data collection and the first explorative research on the AICO Multimodal Corpus. The corpus contains eye-gaze, Kinect, and video recordings of human-robot and human-human interactions, and was collected to study cooperation, engagement and attention of human participants in task-based as well as in chatty type interactive situations. In particular, the goal was to enable comparison between human-human and human-robot interactions, besides studying multimodal behaviour and attention in the different dialogue activities. The robot partner was a humanoid Nao robot, and it was expected that its agent-like behaviour would render humanrobot interactions similar to human-human interaction but also high-light important differences due to the robot’s limited conversational capabilities. The paper reports on the preliminary studies on the corpus, concerning the participants’ eye-gaze and gesturing behaviours,which were chosen as objective measures to study differences in their multimodal behaviour patterns with a human and a robot partner.
Fully data driven Chatbots for non-goal oriented dialogues are known to suffer from inconsistent behaviour across their turns, stemming from a general difficulty in controlling parameters like their assumed background personality and knowledge of facts. One reason for this is the relative lack of labeled data from which personality consistency and fact usage could be learned together with dialogue behaviour. To address this, we introduce a new labeled dialogue dataset in the domain of movie discussions, where every dialogue is based on pre-specified facts and opinions. We thoroughly validate the collected dialogue for adherence of the participants to their given fact and opinion profile, and find that the general quality in this respect is high. This process also gives us an additional layer of annotation that is potentially useful for training models. We introduce as a baseline an end-to-end trained self-attention decoder model trained on this data and show that it is able to generate opinionated responses that are judged to be natural and knowledgeable and show attentiveness.
Data-driven approaches for creating virtual patient dialogue systems require the availability of large data specific to the language,domain and clinical cases studied. Based on the lack of dialogue corpora in French for medical education, we propose an annotatedcorpus of dialogues including medical consultation interactions between doctor and patient. In this work, we detail the building processof the proposed dialogue corpus, describe the annotation guidelines and also present the statistics of its contents. We then conducted aquestion categorization task to evaluate the benefits of the proposed corpus that is made publicly available.
User attributes provide rich and useful information for user understanding, yet structured and easy-to-use attributes are often sparsely populated. In this paper, we leverage dialogues with conversational agents, which contain strong suggestions of user information, to automatically extract user attributes. Since no existing dataset is available for this purpose, we apply distant supervision to train our proposed two-stage attribute extractor, which surpasses several retrieval and generation baselines on human evaluation. Meanwhile, we discuss potential applications (e.g., personalized recommendation and dialogue systems) of such extracted user attributes, and point out current limitations to cast light on future work.
Our goal is to develop an intelligent assistant to support users explore data via visualizations. We have collected a new corpus of conversations, CHICAGO-CRIME-VIS, geared towards supporting data visualization exploration, and we have annotated it for a variety of features, including contextualized dialogue acts. In this paper, we describe our strategies and their evaluation for dialogue act classification. We highlight how thinking aloud affects interpretation of dialogue acts in our setting and how to best capture that information. A key component of our strategy is data augmentation as applied to the training data, since our corpus is inherently small. We ran experiments with the Balanced Bagging Classifier (BAGC), Condiontal Random Field (CRF), and several Long Short Term Memory (LSTM) networks, and found that all of them improved compared to the baseline (e.g., without the data augmentation pipeline). CRF outperformed the other classification algorithms, with the LSTM networks showing modest improvement, even after obtaining a performance boost from domain-trained word embeddings. This result is of note because training a CRF is far less resource-intensive than training deep learning models, hence given a similar if not better performance, traditional methods may still be preferable in order to lower resource consumption.
This paper presents a multimodal corpus of 209 spoken game dialogues between a human and a remote-controlled artificial agent. The interactions involve people collaborating with the agent to identify countries on the world map as quickly as possible, which allows studying rapid and spontaneous dialogue with complex anaphoras, disfluent utterances and incorrect descriptions. The corpus consists of two parts: 8 hours of game interactions have been collected with a virtual unembodied agent online and 26.8 hours have been recorded with a physically embodied robot in a research lab. In addition to spoken audio recordings available for both parts, camera recordings and skeleton-, facial expression- and eye-gaze tracking data have been collected for the lab-based part of the corpus. In this paper, we introduce the pedagogical reference resolution game (RDG-Map) and the characteristics of the corpus collected. We also present an annotation scheme we developed in order to study the dialogue strategies utilized by the players. Based on a subset of 330 minutes of interactions annotated so far, we discuss initial insights into these strategies as well as the potential of the corpus for future research.
A dialogue dataset is an indispensable resource for building a dialogue system. Additional information like emotions and interpersonal relationships labeled on conversations enables the system to capture the emotion flow of the participants in the dialogue. However, there is no publicly available Chinese dialogue dataset with emotion and relation labels. In this paper, we collect the conversions from TV series scripts, and annotate emotion and interpersonal relationship labels on each utterance. This dataset contains 25,548 utterances from 4,142 dialogues. We also set up some experiments to observe the effects of the responded utterance on the current utterance, and the correlation between emotion and relation types in emotion and relation classification tasks.
Datasets featuring modern voice assistants such as Alexa, Siri, Cortana and others allow an easy study of human-machine interactions. But data collections offering an unconstrained, unscripted public interaction are quite rare. Many studies so far have focused on private usage, short pre-defined task or specific domains. This contribution presents a dataset providing a large amount of unconstrained public interactions with a voice assistant. Up to now around 40 hours of device directed utterances were collected during a science exhibition touring through Germany. The data recording was part of an exhibit that engages visitors to interact with a commercial voice assistant system (Amazon’s ALEXA), but did not restrict them to a specific topic. A specifically developed quiz was starting point of the conversation, as the voice assistant was presented to the visitors as a possible joker for the quiz. But the visitors were not forced to solve the quiz with the help of the voice assistant and thus many visitors had an open conversation. The provided dataset – Voice Assistant Conversations in the wild (VACW) – includes the transcripts of both visitors requests and Alexa answers, identified topics and sessions as well as acoustic characteristics automatically extractable from the visitors’ audio files.
The recognition of emotion and dialogue acts enriches conversational analysis and help to build natural dialogue systems. Emotion interpretation makes us understand feelings and dialogue acts reflect the intentions and performative functions in the utterances. However, most of the textual and multi-modal conversational emotion corpora contain only emotion labels but not dialogue acts. To address this problem, we propose to use a pool of various recurrent neural models trained on a dialogue act corpus, with and without context. These neural models annotate the emotion corpora with dialogue act labels, and an ensemble annotator extracts the final dialogue act label. We annotated two accessible multi-modal emotion corpora: IEMOCAP and MELD. We analyzed the co-occurrence of emotion and dialogue act labels and discovered specific relations. For example, Accept/Agree dialogue acts often occur with the Joy emotion, Apology with Sadness, and Thanking with Joy. We make the Emotional Dialogue Acts (EDA) corpus publicly available to the research community for further study and analysis.
PAC0 is a French audio-video conversational corpus made of 15 face-to-face dyadic interactions, lasting around 20 min each. This compared corpus has been created in order to explore the impact of the lack of personal common ground (Clark, 1996) on participants collaboration during conversation and specifically on their smile during topic transitions. We have constituted this conversational corpus " PACO” by replicating the experimental protocol of “Cheese!” (Priego-valverde & al.,2018). The only difference that distinguishes these two corpora is the degree of CG of the interlocutors: in Cheese! interlocutors are friends, while in PACO they do not know each other. This experimental protocol allows to analyze how the participants are getting acquainted. This study brings two main contributions. First, the PACO conversational corpus enables to compare the impact of the interlocutors’ common ground. Second, the semi-automatic smile annotation protocol allows to obtain reliable and reproducible smile annotations while reducing the annotation time by a factor 10. Keywords : Common ground, spontaneous interaction, smile, automatic detection.
This paper deals with the annotation of dialogue acts in a multimodal corpus of first encounter dialogues, i.e. face-to- face dialogues in which two people who meet for the first time talk with no particular purpose other than just talking. More specifically, we describe the method used to annotate dialogue acts in the corpus, including the evaluation of the annotations. Then, we present descriptive statistics of the annotation, particularly focusing on which dialogue acts often follow each other across speakers and which dialogue acts overlap with gestural behaviour. Finally, we discuss how feedback is expressed in the corpus by means of feedback dialogue acts with or without co-occurring gestural behaviour, i.e. multimodal vs. unimodal feedback.
In this study, we propose a conversation-analytic annotation scheme for turn-taking behavior in multi-party conversations. The annotation scheme is motivated by a proposal of a proper model of turn-taking incorporating various ideas developed in the literature of conversation analysis. Our annotation consists of two sets of tags: the beginning and the ending type of the utterance. Focusing on the ending-type tags, in some cases combined with the beginning-type tags, we emphasize the importance of the distinction among four selection types: i) selecting other participant as next speaker, ii) not selecting next speaker but followed by a switch of the speakership, iii) not selecting next speaker and followed by a continuation of the speakership, and iv)being inside a multi-unit turn. Based on the annotation of Japanese multi-party conversations, we analyze how syntactic and prosodic features of utterances vary across the four selection types. The results show that the above four-way distinction is essential to account for the distributions of the syntactic and prosodic features, suggesting the insufficiency of previous turn-taking models that do not consider the distinction between i) and ii) or between ii) or iii).
A dialog system that can monitor the health status of seniors has a huge potential for solving the labor force shortage in the caregiving industry in aging societies. As a part of efforts to create such a system, we are developing two modules that are aimed to correctly interpret user utterances: (i) a yes/no response classifier, which categorizes responses to health-related yes/no questions that the system asks; and (ii) an entailment recognizer, which detects users’ voluntary mentions about their health status. To apply machine learning approaches to the development of the modules, we created large annotated datasets of 280,467 question-response pairs and 38,868 voluntary utterances. For question-response pairs, we asked annotators to avoid direct “yes” or “no” answers, so that our data could cover a wide range of possible natural language responses. The two modules were implemented by fine-tuning a BERT model, which is a recent successful neural network model. For the yes/no response classifier, the macro-average of the average precisions (APs) over all of our four categories (Yes/No/Unknown/Other) was 82.6% (96.3% for “yes” responses and 91.8% for “no” responses), while for the entailment recognizer it was 89.9%.
Interactive dialogue agents like smart speakers have become more and more popular in recent years. These agents are being developed on machine learning technologies that use huge amounts of language resources. However, many entities in specialized fields are struggling to develop their own interactive agents due to a lack of language resources such as dialogue corpora, especially when the end users need interactive agents that offer multilingual support. Therefore, we aim at providing a general design framework for multilingual interactive agents in specialized domains that, it is assumed, have small or non-existent dialogue corpora. To achieve our goal, we first integrate and customize external language services for supporting multilingual functions of interactive agents. Then, we realize context-aware dialogue generation under the situation of small corpora. Third, we develop a gradual design process for acquiring dialogue corpora and improving the interactive agents. We implement a multilingual interactive agent in the field of healthcare and conduct experiments to illustrate the effectiveness of the implemented agent.
In this paper we present investigation of real-life, bi-directional conversations. We introduce the multimodal corpus derived from these natural conversations alternating between human-human and human-robot interactions. The human-robot interactions were used as a control condition for the social nature of the human-human conversations. The experimental set up consisted of conversations between the participant in a functional magnetic resonance imaging (fMRI) scanner and a human confederate or conversational robot outside the scanner room, connected via bidirectional audio and unidirectional videoconferencing (from the outside to inside the scanner). A cover story provided a framework for natural, real-life conversations about images of an advertisement campaign. During the conversations we collected a multimodal corpus for a comprehensive characterization of bi-directional conversations. In this paper we introduce this multimodal corpus which includes neural data from functional magnetic resonance imaging (fMRI), physiological data (blood flow pulse and respiration), transcribed conversational data, as well as face and eye-tracking recordings. Thus, we present a unique corpus to study human conversations including neural, physiological and behavioral data.
This paper presents an original dataset of controlled interactions, focusing on the study of feedback items. It consists on recordings of different conversations between a doctor and a patient, played by actors. In this corpus, the patient is mainly a listener and produces different feedbacks, some of them being (voluntary) incongruent. Moreover, these conversations have been re-synthesized in a virtual reality context, in which the patient is played by an artificial agent. The final corpus is made of different movies of human-human conversations plus the same conversations replayed in a human-machine context, resulting in the first human-human/human-machine parallel corpus. The corpus is then enriched with different multimodal annotations at the verbal and non-verbal levels. Moreover, and this is the first dataset of this type, we have designed an experiment during which different participants had to watch the movies and give an evaluation of the interaction. During this task, we recorded participant’s brain signal. The Brain-IHM dataset is then conceived with a triple purpose: 1/ studying feedbacks by comparing congruent vs. incongruent feedbacks 2/ comparing human-human and human-machine production of feedbacks 3/ studying the brain basis of feedback perception.
This paper describes a schema that enriches Abstract Meaning Representation (AMR) in order to provide a semantic representation for facilitating Natural Language Understanding (NLU) in dialogue systems. AMR offers a valuable level of abstraction of the propositional content of an utterance; however, it does not capture the illocutionary force or speaker’s intended contribution in the broader dialogue context (e.g., make a request or ask a question), nor does it capture tense or aspect. We explore dialogue in the domain of human-robot interaction, where a conversational robot is engaged in search and navigation tasks with a human partner. To address the limitations of standard AMR, we develop an inventory of speech acts suitable for our domain, and present “Dialogue-AMR”, an enhanced AMR that represents not only the content of an utterance, but the illocutionary force behind it, as well as tense and aspect. To showcase the coverage of the schema, we use both manual and automatic methods to construct the “DialAMR” corpus—a corpus of human-robot dialogue annotated with standard AMR and our enriched Dialogue-AMR schema. Our automated methods can be used to incorporate AMR into a larger NLU pipeline supporting human-robot dialogue.
Nowadays, spoken dialogue agents such as communication robots and smart speakers listen to narratives of humans. In order for such an agent to be recognized as a listener of narratives and convey the attitude of attentive listening, it is necessary to generate responsive utterances. Moreover, responsive utterances can express empathy to narratives and showing an appropriate degree of empathy to narratives is significant for enhancing speaker’s motivation. The degree of empathy shown by responsive utterances is thought to depend on their type. However, the relation between responsive utterances and degrees of the empathy has not been explored yet. This paper describes the classification of responsive utterances based on the degree of empathy in order to explain that relation. In this research, responsive utterances are classified into five levels based on the effect of utterances and literature on attentive listening. Quantitative evaluations using 37,995 responsive utterances showed the appropriateness of the proposed classification.
Learning to interview patients to find out their disease is an essential part of the training of medical students. The practical part of this training has traditionally relied on paid actors that play the role of a patient to be interviewed. This process is expensive and severely limits the amount of practice per student. In this work, we present a novel data set and methods based on Natural Language Processing, for making progress towards modern applications and e-learning tools that support this training by providing language-based user interfaces with virtual patients. A data set of german transcriptions from live doctor-patient interviews was collected. These transcriptions are based on audio recordings of exercise sessions within the university and only the doctor’s utterances could be transcribed. We annotated each utterance with an intent inventory characterizing the purpose of the question or statement. For some intent classes, the data only contains a few samples, and we apply Information Retrieval and Deep Learning methods that are robust with respect to small amounts of training data for recognizing the intent of an utterance and providing the correct response. Our results show that the models are effective and they provide baseline performance scores on the data set for further research.
In this paper, we present a tool allowing dynamic prediction and visualization of an individual’s local brain activity during a conversation. The prediction module of this tool is based on classifiers trained using a corpus of human-human and human-robot conversations including fMRI recordings. More precisely, the module takes as input behavioral features computed from raw data, mainly the participant and the interlocutor speech but also the participant’s visual input and eye movements. The visualisation module shows in real-time the dynamics of brain active areas synchronised with the behavioral raw data. In addition, it shows which integrated behavioral features are used to predict the activity in individual brain areas.
In the last years, the state of the art of NLP research has made a huge step forward. Since the release of ELMo (Peters et al., 2018), a new race for the leading scoreboards of all the main linguistic tasks has begun. Several models have been published achieving promising results in all the major NLP applications, from question answering to text classification, passing through named entity recognition. These great research discoveries coincide with an increasing trend for voice-based technologies in the customer care market. One of the next biggest challenges in this scenario will be the handling of multi-turn conversations, a type of conversations that differs from single-turn by the presence of multiple related interactions. The proposed work is an attempt to exploit one of these new milestones to handle multi-turn conversations. MTSI-BERT is a BERT-based model achieving promising results in intent classification, knowledge base action prediction and end of dialogue session detection, to determine the right moment to fulfill the user request. The study about the realization of PuffBot, an intelligent chatbot to support and monitor people suffering from asthma, shows how this type of technique could be an important piece in the development of future chatbots.
We collected a corpus of dialogues in a Wizard of Oz (WOz) setting in the Internet of Things (IoT) domain. We asked users participating in these dialogues to rate the system on a number of aspects, namely, intelligence, naturalness, personality, friendliness, their enjoyment, overall quality, and whether they would recommend the system to others. Then we asked dialogue observers, i.e., Amazon Mechanical Turkers (MTurkers), to rate these dialogues on the same aspects. We also generated simulated dialogues between dialogue policies and simulated users and asked MTurkers to rate them again on the same aspects. Using linear regression, we developed dialogue evaluation functions based on features from the simulated dialogues and the MTurkers’ ratings, the WOz dialogues and the MTurkers’ ratings, and the WOz dialogues and the WOz participants’ ratings. We applied all these dialogue evaluation functions to a held-out portion of our WOz dialogues, and we report results on the predictive power of these different types of dialogue evaluation functions. Our results suggest that for three conversational aspects (intelligence, naturalness, overall quality) just training evaluation functions on simulated data could be sufficient.
We compare two models for corpus-based selection of dialogue responses: one based on cross-language relevance with a cross-language LSTM model. Each model is tested on multiple corpora, collected from two different types of dialogue source material. Results show that while the LSTM model performs adequately on a very large corpus (millions of utterances), its performance is dominated by the cross-language relevance model for a more moderate-sized corpus (ten thousands of utterances).
In this paper, we introduce a multimodal dataset in which subjects are instructing each other how to assemble IKEA furniture. Using the concept of ‘Chinese Whispers’, an old children’s game, we employ a novel method to avoid implicit experimenter biases. We let subjects instruct each other on the nature of the task: the process of the furniture assembly. Uncertainty, hesitations, repairs and self-corrections are naturally introduced in the incremental process of establishing common ground. The corpus consists of 34 interactions, where each subject first assembles and then instructs. We collected speech, eye-gaze, pointing gestures, and object movements, as well as subjective interpretations of mutual understanding, collaboration and task recall. The corpus is of particular interest to researchers who are interested in multimodal signals in situated dialogue, especially in referential communication and the process of language grounding.
The problem of building a coherent and non-monotonous conversational agent with proper discourse and coverage is still an area of open research. Current architectures only take care of semantic and contextual information for a given query and fail to completely account for syntactic and external knowledge which are crucial for generating responses in a chit-chat system. To overcome this problem, we propose an end to end multi-stream deep learning architecture that learns unified embeddings for query-response pairs by leveraging contextual information from memory networks and syntactic information by incorporating Graph Convolution Networks (GCN) over their dependency parse. A stream of this network also utilizes transfer learning by pre-training a bidirectional transformer to extract semantic representation for each input sentence and incorporates external knowledge through the neighborhood of the entities from a Knowledge Base (KB). We benchmark these embeddings on the next sentence prediction task and significantly improve upon the existing techniques. Furthermore, we use AMUSED to represent query and responses along with its context to develop a retrieval based conversational agent which has been validated by expert linguists to have comprehensive engagement with humans.
This paper introduces an approach for annotating eye gaze considering both its social and the referential functions in multi-modal human-human dialogue. Detecting and interpreting the temporal patterns of gaze behavior cues is natural for humans and also mostly an unconscious process. However, these cues are difficult for conversational agents such as robots or avatars to process or generate. The key factor is to recognize these variants and carry out a successful conversation, as misinterpretation can lead to total failure of the given interaction. This paper introduces an annotation scheme for eye-gaze in human-human dyadic interactions that is intended to facilitate the learning of eye-gaze patterns in multi-modal natural dialogue.
We outline the issues and decisions involved in creating a Penn-style treebank of Middle Low German (MLG, 1200-1650), which will form part of the Corpus of Historical Low German (CHLG). The attestation for MLG is rich, but the syntax of the language remains relatively understudied. The development of a syntactically annotated corpus for the language will facilitate future studies with a strong empirical basis, building on recent work which indicates that, syntactically, MLG occupies a position in its own right within West Germanic. In this paper, we describe the background for the corpus and the process by which texts were selected to be included. In particular, we focus on the decisions involved in the syntactic annotation of the corpus, specifically, the practical and linguistic reasons for adopting the Penn annotation scheme, the stages of the annotation process itself, and how we have adapted the Penn scheme for syntactic features specific to MLG. We also discuss the issue of data uncertainty, which is a major issue when building a corpus of an under-researched language stage like MLG, and some novel ways in which we capture this uncertainty in the annotation.
The Book of Hours was the bestseller of the late Middle Ages and Renaissance. It is a historical invaluable treasure, documenting the devotional practices of Christians in the late Middle Ages. Up to now, its textual content has been scarcely studied because of its manuscript nature, its length and its complex content. At first glance, it looks too standardized. However, the study of book of hours raises important challenges: (i) in image analysis, its often lavish ornamentation (illegible painted initials, line-fillers, etc.), abbreviated words, multilingualism are difficult to address in Handwritten Text Recognition (HTR); (ii) its hierarchical entangled structure offers a new field of investigation for text segmentation; (iii) in digital humanities, its textual content gives opportunities for historical analysis. In this paper, we provide the first corpus of books of hours, which consists of Latin transcriptions of 300 books of hours generated by Handwritten Text Recognition (HTR) - that is like Optical Character Recognition (OCR) but for handwritten and not printed texts. We designed a structural scheme of the book of hours and annotated manually two books of hours according to this scheme. Lastly, we performed a systematic evaluation of the main state of the art text segmentation approaches.
Chinese dynastic histories form a large continuous linguistic space of approximately 2000 years, from the 3rd century BCE to the 18th century CE. The histories are documented in Classical (Literary) Chinese in a corpus of over 20 million characters, suitable for the computational analysis of historical lexicon and semantic change. However, there is no freely available open-source corpus of these histories, making Classical Chinese low-resource. This project introduces a new open-source corpus of twenty-four dynastic histories covered by Creative Commons license. An original list of Classical Chinese gender-specific terms was developed as a case study for analyzing the historical linguistic use of male and female terms. The study demonstrates considerable stability in the usage of these terms, with dominance of male terms. Exploration of word meanings uses keyword analysis of focus corpora created for gender-specific terms. This method yields meaningful semantic representations that can be used for future studies of diachronic semantics.
We present a new, extended version of the Royal Society Corpus (RSC), a diachronic corpus of scientific English now covering 300+ years of scientific writing (1665--1996). The corpus comprises 47 837 texts, primarily scientific articles, and is based on publications of the Royal Society of London, mainly its Philosophical Transactions and Proceedings. The corpus has been built on the basis of the FAIR principles and is freely available under a Creative Commons license, excluding copy-righted parts. We provide information on how the corpus can be found, the file formats available for download as well as accessibility via a web-based corpus query platform. We show a number of analytic tools that we have implemented for better usability and provide an example of use of the corpus for linguistic analysis as well as examples of subsequent, external uses of earlier releases. We place the RSC against the background of existing English diachronic/scientific corpora, elaborating on its value for linguistic and humanistic study.
This article presents corpus REDEWIEDERGABE, a German-language historical corpus with detailed annotations for speech, thought and writing representation (ST&WR). With approximately 490,000 tokens, it is the largest resource of its kind. It can be used to answer literary and linguistic research questions and serve as training material for machine learning. This paper describes the composition of the corpus and the annotation structure, discusses some methodological decisions and gives basic statistics about the forms of ST&WR found in this corpus.
In recent years the interest in the use of repositories of literary works has been successful. While many efforts related to Linked Open Data go in the right direction, the use of these repositories for the creation of text corpora enriched with metadata remains difficult and cumbersome. In fact, many of these repositories can be useful to the community not only for the automatic creation of textual corpora but also for retrieving crucial meta-information about texts. In particular, the use of metadata provides the reader with a wealth of information that is often not identifiable in the texts themselves. Our project aims to fill both the access to the textual resources available on the web and the possibility of combining these resources with sources of metadata that can enrich the texts with useful information lengthening the life and maintenance of the data itself. We introduce here a user-friendly web interface of the Digital Humanities toolkit named WeDH with which the user can leverage the encyclopedic knowledge provided by DBpedia, wikidata and VIAF in order to enrich the corpora with bibliographical and exegetical knowledge. WeDH is a collaborative project and we invite anyone who has ideas or suggestions regarding this procedure to reach out to us.
Obituaries contain information about people’s values across times and cultures, which makes them a useful resource for exploring cultural history. They are typically structured similarly, with sections corresponding to Personal Information, Biographical Sketch, Characteristics, Family, Gratitude, Tribute, Funeral Information and Other aspects of the person. To make this information available for further studies, we propose a statistical model which recognizes these sections. To achieve that, we collect a corpus of 20058 English obituaries from TheDaily Item, Remembering.CA and The London Free Press. The evaluation of our annotation guidelines with three annotators on 1008 obituaries shows a substantial agreement of Fleiss κ = 0.87. Formulated as an automatic segmentation task, a convolutional neural network outperforms bag-of-words and embedding-based BiLSTMs and BiLSTM-CRFs with a micro F1 = 0.81.
We describe a new corpus, SLäNDa, the Swedish Literary corpus of Narrative and Dialogue. It contains Swedish literary fiction, which has been manually annotated for cited materials, with a focus on dialogue. The annotation covers excerpts from eight Swedish novels written between 1879–1940, a period of modernization of the Swedish language. SLäNDa contains annotations for all cited materials that are separate from the main narrative, like quotations and signs. The main focus is on dialogue, for which we annotate speech segments, speech tags, and speakers. In this paper we describe the annotation protocol and procedure and show that we can reach a high inter-annotator agreement. In total, SLäNDa contains annotations of 44 chapters with over 220K tokens. The annotation identified 4,733 instances of cited material and 1,143 named speaker–speech mappings. The corpus is useful for developing computational tools for different types of analysis of literary narrative and speech. We perform a small pilot study where we show how our annotation can help in analyzing language change in Swedish. We find that a number of common function words have their modern version appear earlier in speech than in narrative.
We introduce RiQuA (RIch QUotation Annotations), a corpus that provides quotations, including their interpersonal structure (speakers and addressees) for English literary text. The corpus comprises 11 works of 19th-century literature that were manually doubly annotated for direct and indirect quotations. For each quotation, its span, speaker, addressee, and cue are identified (if present). This provides a rich view of dialogue structures not available from other available corpora. We detail the process of creating this dataset, discuss the annotation guidelines, and analyze the resulting corpus in terms of inter-annotator agreement and its properties. RiQuA, along with its annotations guidelines and associated scripts, are publicly available for use, modification, and experimentation.
Song lyrics can be considered as a text genre that has features of both written and spoken discourse, and potentially provides extensive linguistic and cultural information to scientists from various disciplines. However, pop songs play a rather subordinate role in empirical language research so far - most likely due to the absence of scientifically valid and sustainable resources. The present paper introduces a multiply annotated corpus of German lyrics as a publicly available basis for multidisciplinary research. The resource contains three types of data for the investigation and evaluation of quite distinct phenomena: TEI-compliant song lyrics as primary data, linguistically and literary motivated annotations, and extralinguistic metadata. It promotes empirically/statistically grounded analyses of genre-specific features, systemic-structural correlations and tendencies in the texts of contemporary pop music. The corpus has been stratified into thematic and author-specific archives; the paper presents some basic descriptive statistics, as well as the public online frontend with its built-in evaluation forms and live visualisations.
This paper presents the BDCamões Collection of Portuguese Literary Documents, a new corpus of literary texts written in Portuguese that in its inaugural version includes close to 4 million words from over 200 complete documents from 83 authors in 14 genres, covering a time span from the 16th to the 21st century, and adhering to different orthographic conventions. Many of the texts in the corpus have also been automatically parsed with state-of-the-art language processing tools, forming the BDCamões Treebank subcorpus. This set of characteristics makes of BDCamões an invaluable resource for research in language technology (e.g. authorship detection, genre classification, etc.) and in language science and digital humanities (e.g. comparative literature, diachronic linguistics, etc.).
Languages change over time and, thanks to the abundance of digital corpora, their evolutionary analysis using computational techniques has recently gained much research attention. In this paper, we focus on creating a dataset to support investigating the similarity in evolution between different languages. We look in particular into the similarities and differences between the use of corresponding words across time in English and French, two languages from different linguistic families yet with shared syntax and close contact. For this we select a set of cognates in both languages and study their frequency changes and correlations over time. We propose a new dataset for computational approaches of synchronized diachronic investigation of language pairs, and subsequently show novel findings stemming from the cognate-focused diachronic comparison of the two chosen languages. To the best of our knowledge, the present study is the first in the literature to use computational approaches and large data to make a cross-language diachronic analysis.
We present NordiCon, a database containing medieval Nordic personal names attested in Continental sources. The database combines formally interpreted and richly interlinked onomastic data with digitized versions of the medieval manuscripts from which the data originate and information on the tokens’ context. The structure of NordiCon is inspired by other online historical given name dictionaries. It takes up challenges reported on in previous works, such as how to cover material properties of a name token and how to define lemmatization principles, and elaborates on possible solutions. The lemmatization principles for NordiCon are further developed in order to facilitate the connection to other name dictionaries and corpuses, and the integration of the database into Språkbanken Text, an infrastructure containing modern and historical written data.
Google Scholar is the largest web search engine for academic literature that also provides access to rich metadata associated with the papers. The ACL Anthology (AA) is the largest repository of articles on Natural Language Processing (NLP). We extracted information from AA for about 44 thousand NLP papers and identified authors who published at least three papers there. We then extracted citation information from Google Scholar for all their papers (not just their AA papers). This resulted in a dataset of 1.1 million papers and associated Google Scholar information. We aligned the information in the AA and Google Scholar datasets to create the NLP Scholar Dataset – a single unified source of information (from both AA and Google Scholar) for tens of thousands of NLP papers. It can be used to identify broad trends in productivity, focus, and impact of NLP research. We present here initial work on analyzing the volume of research in NLP over the years and identifying the most cited papers in NLP. We also list a number of additional potential applications.
There exist as many as 7000 natural languages in the world, and a huge number of documents describing those languages have been produced over the years. Most of those documents are in paper format. Any attempts to use modern computational techniques and tools to process those documents will require them to be digitized first. In this paper, we report a multilingual digitized version of thousands of such documents searchable through some well-established corpus infrastructures. The corpus is annotated with various meta, word, and text level attributes to make searching and analysis easier and more useful.
We propose a mixed methods approach to the identification of scribes and authors in handwritten documents, and present LiViTo, a software tool which combines linguistic insights and computer vision techniques in order to assist researchers in the analysis of handwritten historical documents. Our research shows that it is feasible to train neural networks for the automatic transcription of handwritten documents and to use these transcriptions as input for further learning processes. Hypotheses about scribes can be tested effectively by extracting visual handwriting features and clustering them appropriately. Methods from linguistics and from computer vision research integrate into a mixed methods system, with benefits on both sides. LiViTo was trained with historical Czech texts by 18th century immigrants to Berlin, a total of 564 pages from a corpus of about 5000 handwritten pages without indication of author or scribe. We provide an overview of the three-year development of LiViTo and an introduction into its methodology and its functions. We then present our findings concerning the corpus of Berlin Czech manuscripts and discuss possible further usage scenarios.
The annotation of texts and other material in the field of digital humanities and Natural Language Processing (NLP) is a common task of research projects. At the same time, the annotation of corpora is certainly the most time- and cost-intensive component in research projects and often requires a high level of expertise according to the research interest. However, for the annotation of texts, a wide range of tools is available, both for automatic and manual annotation. Since the automatic pre-processing methods are not error-free and there is an increasing demand for the generation of training data, also with regard to machine learning, suitable annotation tools are required. This paper defines criteria of flexibility and efficiency of complex annotations for the assessment of existing annotation tools. To extend this list of tools, the paper describes TextAnnotator, a browser-based, multi-annotation system, which has been developed to perform platform-independent multimodal annotations and annotate complex textual structures. The paper illustrates the current state of development of TextAnnotator and demonstrates its ability to evaluate annotation quality (inter-annotator agreement) at runtime. In addition, it will be shown how annotations of different users can be performed simultaneously and collaboratively on the same document from different platforms using UIMA as the basis for annotation.
Deduplication is the task of identifying near and exact duplicate data items in a collection. In this paper, we present a novel method for deduplication of scholarly documents. We develop a hybrid model which uses structural similarity (locality sensitive hashing) and meaning representation (word embeddings) of document texts to determine (near) duplicates. Our collection constitutes a subset of multidisciplinary scholarly documents aggregated from research repositories. We identify several issues causing data inaccuracies in such collections and motivate the need for deduplication. In lack of existing dataset suitable for study of deduplication of scholarly documents, we create a ground truth dataset of 100K scholarly documents and conduct a series of experiments to empirically establish optimal values for the parameters of our deduplication method. Experimental evaluation shows that our method achieves a macro F1-score of 0.90. We productionise our method as a publicly accessible web API service serving deduplication of scholarly documents in real time.
“Voices of the Great War” is the first large corpus of Italian historical texts dating back to the period of First World War. This corpus differs from other existing resources in several respects. First, from the linguistic point of view it gives account of the wide range of varieties in which Italian was articulated in that period, namely from a diastratic (educated vs. uneducated writers), diaphasic (low/informal vs. high/formal registers) and diatopic (regional varieties, dialects) points of view. From the historical perspective, through a collection of texts belonging to different genres it represents different views on the war and the various styles of narrating war events and experiences. The final corpus is balanced along various dimensions, corresponding to the textual genre, the language variety used, the author type and the typology of conveyed contents. The corpus is fully annotated with lemmas, part-of-speech, terminology, and named entities. Significant corpus samples representative of the different “voices” have also been enriched with meta-linguistic and syntactic information. The layer of syntactic annotation forms the first nucleus of an Italian historical treebank complying with the Universal Dependencies standard. The paper illustrates the final resource, the methodology and tools used to build it, and the Web Interface for navigating it.
DEbateNet-migr15 is a manually annotated dataset for German which covers the public debate on immigration in 2015. The building block of our annotation is the political science notion of a claim, i.e., a statement made by a political actor (a politician, a party, or a group of citizens) that a specific action should be taken (e.g., vacant flats should be assigned to refugees). We identify claims in newspaper articles, assign them to actors and fine-grained categories and annotate their polarity and date. The aim of this paper is two-fold: first, we release the full DEbateNet-mig15 corpus and document it by means of a quantitative and qualitative analysis; second, we demonstrate its application in a discourse network analysis framework, which enables us to capture the temporal dynamics of the political debate
This paper documents a corpus of political speeches in Spanish. The documents in the corpus belong to the Christmas speeches that have been delivered yearly by the head of state of Spain since 1937. The historical period covered by these speeches ranges from the Spanish Civil War and the Francoist dictatorship up until today. As a result, the corpus reflects some of the most significant events and political changes in the recent history of Spain. Up until now, the speeches as a whole had not been collected into a single, systematic and reusable resource, as most of the texts were scattered among different sources. The paper describes: (1) the composition of the corpus; (2) the Python interface that facilitates querying and analyzing the corpus using the NLTK and spaCy libraries and (3) a set of HTML visualizations aimed at the general public to navigate the corpus and explore differences between TF-IDF frequencies.
The present work introduces a new Latin treebank that follows the Universal Dependencies (UD) annotation standard. The treebank is obtained from the automated conversion of the Late Latin Charter Treebank 2 (LLCT2), originally in the Prague Dependency Treebank (PDT) style. As this treebank consists of Early Medieval legal documents, its language variety differs considerably from both the Classical and Medieval learned varieties prevalent in the other currently available UD Latin treebanks. Consequently, besides significant phenomena from the perspective of diachronic linguistics, this treebank also poses several challenging technical issues for the current and future syntactic annotation of Latin in the UD framework. Some of the most relevant cases are discussed in depth, with comparisons between the original PDT and the resulting UD annotations. Additionally, an overview of the UD-style structure of the treebank is given, and some diachronic aspects of the transition from Latin to Romance languages are highlighted.
In order to access indigenous, regional knowledge contained in language corpora, semantic tools and network methods are most typically employed. In this paper we present an approach for the identification of dialectal variations of words, or words that do not pertain to High German, on the example of non-standard language legacy collection questionnaires of the Bavarian Dialects in Austria (DBÖ). Based on selected cultural categories relevant to the wider project context, common words from each of these cultural categories and their lemmas using GermaLemma were identified. Through word embedding models the semantic vicinity of each word was explored, followed by the use of German Wordnet (Germanet) and the Hunspell tool. Whilst none of these tools have a comprehensive coverage of standard German words, they serve as an indication of dialects in specific semantic hierarchies. Methods and tools applied in this study may serve as an example for other similar projects dealing with non-standard or endangered language collections, aiming to access, analyze and ultimately preserve native regional language heritage.
Yiddish is a low-resource language belonging to the Germanic language family and written using the Hebrew alphabet. As a language, Yiddish can be considered resource-poor as it lacks both public accessible corpora and a widely-used standard orthography, with various countries and organizations influencing the spellings speakers use. While existing corpora of Yiddish text do exist, they are often only written in a single, potentially non-standard orthography, with no parallel version with standard orthography available. In this work, we introduce the first multi-orthography parallel corpus of Yiddish nouns built by scraping word entries from Wiktionary. We also demonstrate how the corpus can be used to bootstrap a transliteration model using the Sequitur-G2P grapheme-to-phoneme conversion toolkit to map between various orthographies. Our trained system achieves error rates between 16.79% and 28.47% on the test set, depending on the orthographies considered. In addition to quantitative analysis, we also conduct qualitative error analysis of the trained system, concluding that non-phonetically spelled Hebrew words are the largest cause of error. We conclude with remarks regarding future work and release the corpus and associated code under a permissive license for the larger community to use.
The final outcome of the project Open Access Database: Adjective-Adverb Interfaces in Romance is an annotated and lemmatised corpus of various linguistic phenomena related to Romance adjectives with adverbial functions. The data is published under open-access and aims to serve linguistic research based on transparent and accessible corpus-based data. The annotation model was developed to offer a cross-linguistic categorization model for the heterogeneous word-class “adverb”, based on its diverse forms, functions and meanings. The project focuses on the interoperability and accessibility of data, with particular respect to reusability in the sense of the FAIR Data Principles. Topics presented by this paper include data compilation and creation, annotation in XML/TEI, data preservation and publication process by means of the GAMS repository and accessibility via a search interface. These aspects are tied together by semantic technologies, using an ontology-based approach.
Following decades of massive digitization, an unprecedented amount of historical document facsimiles can now be retrieved and accessed via cultural heritage online portals. If this represents a huge step forward in terms of preservation and accessibility, the next fundamental challenge– and real promise of digitization– is to exploit the contents of these digital assets, and therefore to adapt and develop appropriate language technologies to search and retrieve information from this ‘Big Data of the Past’. Yet, the application of text processing tools on historical documents in general, and historical newspapers in particular, poses new challenges, and crucially requires appropriate language resources. In this context, this paper presents a collection of historical newspaper data sets composed of text and image resources, curated and published within the context of the ‘impresso - Media Monitoring of the Past’ project. With corpora, benchmarks, semantic annotations and language models in French, German and Luxembourgish covering ca. 200 years, the objective of the impresso resource collection is to contribute to historical language resources, and thereby strengthen the robustness of approaches to non-standard inputs and foster efficient processing of historical documents.
The massive digitization efforts related to historical newspapers over the past decades have focused on mass media sources and ordinary people as their primary recipients. Much less attention has been paid to newspapers published for a more specialized audience, e.g., those aiming at scholarly or cultural exchange within intellectual communities much narrower in scope, such as newspapers devoted to music criticism, arts or philosophy. Only some few of these specialized newspapers have been digitized up until now, but they are usually not well curated in terms of digitization quality, data formatting, completeness, redundancy (de-duplication), supply of metadata, and, hence, searchability. This paper describes our approach to eliminate these drawbacks for a major German-language newspaper resource of the Romantic Age, the Allgemeine Musikalische Zeitung (General Music Gazette). We here focus on a workflow that copes with a posteriori digitization problems, inconsistent OCRing and index building for searchability. In addition, we provide a user-friendly graphic interface to empower content-centric access to this (and other) digital resource(s) adopting open-source software for the purpose of Web presentation.
The method of stylometry by most frequent words does not allow direct comparison of original texts and their translations, i.e. across languages. For instance, in a bilingual Czech-German text collection containing parallel texts (originals and translations in both directions, along with Czech and German translations from other languages), authors would not cluster across languages, since frequency word lists for any Czech texts are obviously going to be more similar to each other than to a German text, and the other way round. We have tried to come up with an interlingua that would remove the language-specific features and possibly keep the linguistically independent features of individual author signal, if they exist. We have tagged, lemmatized, and parsed each language counterpart with the corresponding language model in UDPipe, which provides a linguistic markup that is cross-lingual to a significant extent. We stripped the output of language-dependent items, but that alone did not help much. As a next step, we transformed the lemmas of both language counterparts into shared pseudolemmas based on a very crude Czech-German glossary, with a 95.6% success. We show that, for stylometric methods based on the most frequent words, we can do without translations.
We present in this work a universal, character-based method for representing sentences so that one can thereby calculate the distance between any two sentence pair. With a small alphabet, it can function as a proxy of phonemes, and as one of its main uses, we carry out dialect clustering: cluster a dialect/sub-language mixed corpus into sub-groups and see if they coincide with the conventional boundaries of dialects and sub-languages. By using data with multiple Japanese dialects and multiple Slavic languages, we report how well each group clusters, in a manner to partially respond to the question of what separates languages from dialects.
Using a model trained to predict discourse markers between sentence pairs, we predict plausible markers between sentence pairs with a known semantic relation (provided by existing classification datasets). These predictions allow us to study the link between discourse markers and the semantic relations annotated in classification datasets. Handcrafted mappings have been proposed between markers and discourse relations on a limited set of markers and a limited set of categories, but there exists hundreds of discourse markers expressing a wide variety of relations, and there is no consensus on the taxonomy of relations between competing discourse theories (which are largely built in a top-down fashion). By using an automatic prediction method over existing semantically annotated datasets, we provide a bottom-up characterization of discourse markers in English. The resulting dataset, named DiscSense, is publicly available.
This paper introduces ThemePro, a toolkit for the automatic analysis of thematic progression. Thematic progression is relevant to natural language processing (NLP) applications dealing, among others, with discourse structure, argumentation structure, natural language generation, summarization and topic detection. A web platform demonstrates the potential of this toolkit and provides a visualization of the results including syntactic trees, hierarchical thematicity over propositions and thematic progression over whole texts.
We introduce a corpus of the 2016 U.S. presidential debates and commentary, containing 4,648 argumentative propositions annotated with fine-grained proposition types. Modern machine learning pipelines for analyzing argument have difficulty distinguishing between types of propositions based on their factuality, rhetorical positioning, and speaker commitment. Inability to properly account for these facets leaves such systems inaccurate in understanding of fine-grained proposition types. In this paper, we demonstrate an approach to annotating for four complex proposition types, namely normative claims, desires, future possibility, and reported speech. We develop a hybrid machine learning and human workflow for annotation that allows for efficient and reliable annotation of complex linguistic phenomena, and demonstrate with preliminary analysis of rhetorical strategies and structure in presidential debates. This new dataset and method can support technical researchers seeking more nuanced representations of argument, as well as argumentation theorists developing new quantitative analyses.
Chinese discourse parsing, which aims to identify the hierarchical relationships of Chinese elementary discourse units, has not yet a consistent evaluation metric. Although Parseval is commonly used, variations of evaluation differ from three aspects: micro vs. macro F1 scores, binary vs. multiway ground truth, and left-heavy vs. right-heavy binarization. In this paper, we first propose a neural network model that unifies a pre-trained transformer and CKY-like algorithm, and then compare it with the previous models with different evaluation scenarios. The experimental results show that our model outperforms the previous systems. We conclude that (1) the pre-trained context embedding provides effective solutions to deal with implicit semantics in Chinese texts, and (2) using multiway ground truth is helpful since different binarization approaches lead to significant differences in performance.
Text corpora annotated with language-related properties are an important resource for the development of Language Technology. The current work contributes a new resource for Chinese Language Technology and for Chinese-English translation, in the form of a set of TED talks (some originally given in English, some in Chinese) that have been annotated with discourse relations in the style of the Penn Discourse TreeBank, adapted to properties of Chinese text that are not present in English. The resource is currently unique in annotating discourse-level properties of planned spoken monologues rather than of written text. An inter-annotator agreement study demonstrates that the annotation scheme is able to achieve highly reliable results.
Although NLP research on argument mining has advanced considerably in recent years, most studies draw on corpora of asynchronous and written texts, often produced by individuals. Few published corpora of synchronous, multi-party argumentation are available. The Discussion Tracker corpus, collected in high school English classes, is an annotated dataset of transcripts of spoken, multi-party argumentation. The corpus consists of 29 multi-party discussions of English literature transcribed from 985 minutes of audio. The transcripts were annotated for three dimensions of collaborative argumentation: argument moves (claims, evidence, and explanations), specificity (low, medium, high) and collaboration (e.g., extensions of and disagreements about others’ ideas). In addition to providing descriptive statistics on the corpus, we provide performance benchmarks and associated code for predicting each dimension separately, illustrate the use of the multiple annotations in the corpus to improve performance via multi-task learning, and finally discuss other ways the corpus might be used to further NLP research.
Shallow Discourse Parsing (SDP), the identification of coherence relations between text spans, relies on large amounts of training data, which so far exists only for English - any other language is in this respect an under-resourced one. For those languages where machine translation from English is available with reasonable quality, MT in conjunction with annotation projection can be an option for producing an SDP resource. In our study, we translate the English Penn Discourse TreeBank into German and experiment with various methods of annotation projection to arrive at the German counterpart of the PDTB. We describe the key characteristics of the corpus as well as some typical sources of errors encountered during its creation. Then we evaluate the GermanPDTB by training components for selected sub-tasks of discourse parsing on this silver data and compare performance to the same components when trained on the gold, original PDTB corpus.
People can extract precise, complex logical meanings from text in documents such as tax forms and game rules, but language processing systems lack adequate training and evaluation resources to do these kinds of tasks reliably. This paper describes a corpus of annotated typed lambda calculus translations for approximately 2,000 sentences in Simple English Wikipedia, which is assumed to constitute a broad-coverage domain for precise, complex descriptions. The corpus described in this paper contains a large number of quantifiers and interesting scoping configurations, and is presented specifically as a resource for quantifier scope disambiguation systems, but also more generally as an object of linguistic study.
We present the Potsdam Commentary Corpus 2.2, a German corpus of news editorials annotated on several different levels. New in the 2.2 version of the corpus are two additional annotation layers for coherence relations following the Penn Discourse TreeBank framework. Specifically, we add relation senses to an already existing layer of discourse connectives and their arguments, and we introduce a new layer with additional coherence relation types, resulting in a German corpus that mirrors the PDTB. The aim of this is to increase usability of the corpus for the task of shallow discourse parsing. In this paper, we provide inter-annotator agreement figures for the new annotations and compare corpus statistics based on the new annotations to the equivalent statistics extracted from the PDTB.
In this paper, we report on our experiments towards the creation of a corpus for coherence evaluation. Most corpora for textual coherence evaluation are composed of randomly shuffled sentences that focus on sentence ordering, regardless of whether the sentences were originally related by a discourse relation. To the best of our knowledge, no publicly available corpus has been designed specifically for the evaluation of coherence of known discursive units. In this paper, we focus on coherence modeling at the intra-discursive level and describe our approach to build a corpus of incoherent pairs of sentences. We experimented with a variety of corruption strategies to create synthetic incoherent pairs of discourse arguments from coherent ones. Using discourse argument pairs from the Penn Discourse Tree Bank, we generate incoherent discourse argument pairs, by swapping either their discourse connective or a discourse argument. To evaluate how incoherent the generated corpora are, we use a convolutional neural network to try to distinguish the original pairs from the corrupted ones. Results of the classifier as well as a manual inspection of the corpora show that generating such corpora is still a challenge as the generated instances are clearly not “incoherent enough”, indicating that more effort should be spent on developing more robust ways of generating incoherent corpora.
This paper describes an accurate framework for carrying out multi-lingual discourse segmentation with BERT (Devlin et al., 2019). The model is trained to identify segments by casting the problem as a token classification problem and jointly learning syntactic features like part-of-speech tags and dependency relations. This leads to significant improvements in performance. Experiments are performed in different languages, such as English, Dutch, German, Portuguese Brazilian and Basque to highlight the cross-lingual effectiveness of the segmenter. In particular, the model achieves a state-of-the-art F-score of 96.7 for the RST-DT corpus (Carlson et al., 2003) improving on the previous best model by 7.2%. Additionally, a qualitative explanation is provided for how proposed changes contribute to model performance by analyzing errors made on the test data.
Gestures are an important component of non–verbal communication. This has an increasing potential in human–computer interaction. For example, Navarretta (2017b) uses sequences of speech and pauses together with co–speech gestures produced by Barack Obama in order to predict audience response, such as applause. The aim of this study is to explore the role of speech pauses and gestures alone as predictors of audience reaction without other types of speech information. For this work, we created a corpus of speeches held by Donald Trump before and during his time as president between 2016 and 2019. The data were transcribed with pause information and co–speech gestures were annotated as well as audience responses. Gestures and long silent pauses of the duration of at least 0.5 seconds are the input of computational models to predict audience reaction. The results of this study indicate that especially head movements and facial expressions play an important role and they confirm that gestures can to some extent be used to predict audience reaction independently of speech.
We introduce the first version of GeCzLex, an online electronic resource for translation equivalents of Czech and German discourse connectives. The lexicon is one of the outcomes of the research on anaphoricity and long-distance relations in discourse, it contains at present anaphoric connectives (ACs) for Czech and German connectives, and further their possible translations documented in bilingual parallel corpora (not necessarily anaphoric). As a basis, we use two existing monolingual lexicons of connectives: the Lexicon of Czech Discourse Connectives (CzeDLex) and the Lexicon of Discourse Markers (DiMLex) for German, interlink their relevant entries via semantic annotation of the connectives (according to the PDTB 3 sense taxonomy) and statistical information of translation possibilities from the Czech and German parallel data of the InterCorp project. The lexicon is, as far as we know, the first bilingual inventory of connectives with linkage on the level of individual entries, and a first attempt to systematically describe devices engaged in long-distance, non-local discourse coherence. The lexicon is freely available under the Creative Commons License.
We present DiMLex-Bangla, a newly developed lexicon of discourse connectives in Bangla. The lexicon, upon completion of its first version, contains 123 Bangla connective entries, which are primarily compiled from the linguistic literature and translation of English discourse connectives. The lexicon compilation is later augmented by adding more connectives from a currently developed corpus, called the Bangla RST Discourse Treebank (Das and Stede, 2018). DiMLex-Bangla provides information on syntactic categories of Bangla connectives, their discourse semantics and non-connective uses (if any). It uses the format of the German connective lexicon DiMLex (Stede and Umbach, 1998), which provides a cross-linguistically applicable XML schema. The resource is the first of its kind in Bangla, and is freely available for use in studies on discourse structure and computational applications.
This paper describes a novel application of semi-supervision for shallow discourse parsing. We use a neural approach for sequence tagging and focus on the extraction of explicit discourse arguments. First, additional unlabeled data is prepared for semi-supervised learning. From this data, weak annotations are generated in a first setting and later used in another setting to study performance differences. In our studies, we show an increase in the performance of our models that ranges between 2-10% F1 score. Further, we give some insights to the generated discourse annotations and compare the developed additional relations with the training relations. We release this new dataset of explicit discourse arguments to enable the training of large statistical models.
This paper presents WikiPossessions, a new benchmark corpus for the task of temporally-oriented possession (TOP), or tracking objects as they change hands over time. We annotate Wikipedia articles for 90 different well-known artifacts paintings, diamonds, and archaeological artifacts), producing 799 artifact-possessor relations with associated attributes. For each article, we also produce a full possession timeline. The full version of the task combines straightforward entity-relation extraction with complex temporal reasoning, as well as verification of textual support for the relevant types of knowledge. Specifically, to complete the full TOP task for a given article, a system must do the following: a) identify possessors; b) anchor possessors to times/events; c) identify temporal relations between each temporal anchor and the possession relation it corresponds to; d) assign certainty scores to each possessor and each temporal relation; and e) assemble individual possession events into a global possession timeline. In addition to the corpus, we release evaluation scripts and a baseline model for the task.
We present a new dataset of TED-talks annotated with the questions they evoke and, where available, the answers to these questions. Evoked questions represent a hitherto mostly unexplored type of linguistic data, which promises to open up important new lines of research, especially related to the Question Under Discussion (QUD)-based approach to discourse structure. In this paper we introduce the method and open the first installment of our data to the public. We summarize and explore the current dataset, illustrate its potential by providing new evidence for the relation between predictability and implicitness – capitalizing on the already existing PDTB-style annotations for the texts we use – and outline its potential for future research. The dataset should be of interest, at its current scale, to researchers on formal and experimental pragmatics, discourse coherence, information structure, discourse expectations and processing. Our data-gathering procedure is designed to scale up, relying on crowdsourcing by non-expert annotators, with its utility for Natural Language Processing in mind (e.g., dialogue systems, conversational question answering).
CzeDLex is an electronic lexicon of Czech discourse connectives with its data coming from a large treebank annotated with discourse relations. Its new version CzeDLex 0.6 (as compared with the previous version 0.5, which was published in 2017) is significantly larger with respect to manually processed entries. Also, its structure has been modified to allow for primary connectives to appear with multiple entries for a single discourse sense. The lexicon comes in several formats, being both human and machine readable, and is available for searching in PML Tree Query, a user-friendly and powerful search tool for all kinds of linguistically annotated treebanks. The main purpose of this paper/demo is to present the new version of the lexicon and to demonstrate possibilities of mining various types of information from the lexicon using PML Tree Query; we present several examples of search queries over the lexicon data along with their results. The new version of the lexicon, CzeDLex~0.6, is available on-line and was officially released in December 2019 under the Creative Commons License.
Persuasions are common in online arguments such as discussion forums. To analyze persuasive strategies, it is important to understand how individuals construct posts and comments based on the semantics of the argumentative components. In addition to understanding how we construct arguments, understanding how a user post interacts with other posts (i.e., argumentative inter-post relation) still remains a challenge. Therefore, in this study, we developed a novel annotation scheme and corpus that capture both user-generated inner-post arguments and inter-post relations between users in ChangeMyView, a persuasive forum. Our corpus consists of arguments with 4612 elementary units (EUs) (i.e., propositions), 2713 EU-to-EU argumentative relations, and 605 inter-post argumentative relations in 115 threads. We analyzed the annotated corpus to identify the characteristics of online persuasive arguments, and the results revealed persuasive documents have more claims than non-persuasive ones and different interaction patterns among persuasive and non-persuasive documents. Our corpus can be used as a resource for analyzing persuasiveness and training an argument mining system to identify and extract argument structures. The annotated corpus and annotation guidelines have been made publicly available.
We present a work aiming to generate adapted content for dyslexic children for French, in the context of the ALECTOR project. Thus, we developed a system to transform the texts at the discourse level. This system modifies the coreference chains, which are markers of text cohesion, by using rules. These rules were designed following a careful study of coreference chains in both original texts and its simplified versions. Moreover, in order to define reliable transformation rules, we analysed several coreference properties as well as the concurrent simplification operations in the aligned texts. This information is coded together with a coreference resolution system and a text rewritten tool in the proposed system, which comprise a coreference module specialised in written text and seven text transformation operations. The evaluation of the system firstly focused on check the simplification by manual validation of three judges. These errors were grouped into five classes that combined can explain 93% of the errors. The second evaluation step consisted of measuring the simplification perception by 23 judges, which allow us to measure the simplification impact of the proposed rules.
BERT, a neural network-based language model pre-trained on large corpora, is a breakthrough in natural language processing, significantly outperforming previous state-of-the-art models in numerous tasks. However, there have been few reports on its application to implicit discourse relation classification, and it is not clear how BERT is best adapted to the task. In this paper, we test three methods of adaptation. (1) We perform additional pre-training on text tailored to discourse classification. (2) In expectation of knowledge transfer from explicit discourse relations to implicit discourse relations, we add a task named explicit connective prediction at the additional pre-training step. (3) To exploit implicit connectives given by treebank annotators, we add a task named implicit connective prediction at the fine-tuning step. We demonstrate that these three techniques can be combined straightforwardly in a single training pipeline. Through comprehensive experiments, we found that the first and second techniques provide additional gain while the last one did not.
This paper focuses on the automatic detection of hidden intentions of speakers in questions asked during meals. Our corpus is composed of a set of transcripts of spontaneous oral conversations from ESLO’s corpora. We suggest a typology of these intentions based on our research work and the exploration and annotation of the corpus, in which we define two “explicit” categories (request for agreement and request for information) and three “implicit” categories (opinion, will and doubt). We implement a supervised automatic classification model based on annotated data and selected linguistic features and we evaluate its results and performances. We finally try to interpret these results by looking more deeply and specifically into the predictions of the algorithm and the features it used. There are many motivations for this work which are part of ongoing challenges such as opinion analysis, irony detection or the development of conversational agents.
Automating straightforward clinical tasks can reduce workload for healthcare professionals, increase accessibility for geographically-isolated patients, and alleviate some of the economic burdens associated with healthcare. A variety of preliminary screening procedures are potentially suitable for automation, and one such domain that has remained underexplored to date is that of structured clinical interviews. A task-specific dialogue agent is needed to automate the collection of conversational speech for further (either manual or automated) analysis, and to build such an agent, a dialogue manager must be trained to respond to patient utterances in a manner similar to a human interviewer. To facilitate the development of such an agent, we propose an annotation schema for assigning dialogue act labels to utterances in patient-interviewer conversations collected as part of a clinically-validated cognitive health screening task. We build a labeled corpus using the schema, and show that it is characterized by high inter-annotator agreement. We establish a benchmark dialogue act classification model for the corpus, thereby providing a proof of concept for the proposed annotation schema. The resulting dialogue act corpus is the first such corpus specifically designed to facilitate automated cognitive health screening, and lays the groundwork for future exploration in this area.
Much research has been done within the social sciences on the interpretation and influence of stigma on human behaviour and health, which result in out-of-group exclusion, distancing, cognitive separation, status loss, discrimination, in-group pressure, and often lead to disengagement, non-adherence to treatment plan, and prescriptions by the doctor. However, little work has been conducted on computational identification of stigma in general and in social media discourse in particular. In this paper, we develop the annotation scheme and improve the annotation process for stigma identification, which can be applied to other health-care domains. The data from pro-vaccination and anti-vaccination discussion groups are annotated by trained annotators who have professional background in social science and health-care studies, therefore the group can be considered experts on the subject in comparison to non-expert crowd. Amazon MTurk annotators is another group of annotator with no knowledge on their education background, they are initially treated as non-expert crowd on the subject matter of stigma. We analyze the annotations with visualisation techniques, features from LIWC (Linguistic Inquiry and Word Count) list and make prediction based on bi-grams with traditional and deep learning models. Data augmentation method and application of CNN show high performance accuracy in comparison to other models. Success of the rigorous annotation process on identifying stigma is reconfirmed by achieving high prediction rate with CNN.
In this paper, we present a new corpus consisting of sentences from Hindi short stories annotated for five different discourse modes argumentative, narrative, descriptive, dialogic and informative. We present a detailed account of the entire data collection and annotation processes. The annotations have a very high inter-annotator agreement (0.87 k-alpha). We analyze the data in terms of label distributions, part of speech tags, and sentence lengths. We characterize the performance of various classification algorithms on this dataset and perform ablation studies to understand the nature of the linguistic models suitable for capturing the nuances of the embedded discourse structures in the presented corpus.
Wikipedia is a great source of general world knowledge which can guide NLP models better understand their motivation to make predictions. Structuring Wikipedia is the initial step towards this goal which can facilitate fine-grain classification of articles. In this work, we introduce the Shinra 5-Language Categorization Dataset (SHINRA-5LDS), a large multi-lingual and multi-labeled set of annotated Wikipedia articles in Japanese, English, French, German, and Farsi using Extended Named Entity (ENE) tag set. We evaluate the dataset using the best models provided for ENE label set classification and show that the currently available classification models struggle with large datasets using fine-grained tag sets.
In this paper, we address the lack of resources for opinion and emotion analysis related to North African dialects, targeting Algerian dialect. We present TWIFIL (TWItter proFILing) a collaborative annotation platform for crowdsourcing annotation of tweets at different levels of granularity. The plateform allowed the creation of the largest Algerian dialect dataset annotated for both sentiment (9,000 tweets), emotion (about 5,000 tweets) and extra-linguistic information including author profiling (age and gender). The annotation resulted also in the creation of the largest Algerien dialect subjectivity lexicon of about 9,000 entries which can constitute a valuable resources for the development of future NLP applications for Algerian dialect. To test the validity of the dataset, a set of deep learning experiments were conducted to classify a given tweet as positive, negative or neutral. We discuss our results and provide an error analysis to better identify classification errors.
Transformer models, trained and publicly released over the last couple of years, have proved effective in many NLP tasks. We wished to test their usefulness in particular on the stance detection task. We performed experiments on the data from the Fake News Challenge Stage 1 (FNC-1). We were indeed able to improve the reported SotA on the challenge, by exploiting the generalization power of large language models based on Transformer architecture. Specifically (1) we improved the FNC-1 best performing model adding BERT sentence embedding of input sequences as a model feature, (2) we fine-tuned BERT, XLNet, and RoBERTa transformers on FNC-1 extended dataset and obtained state-of-the-art results on FNC-1 task.
We introduce a new classification task for scientific statements and release a large-scale dataset for supervised learning. Our resource is derived from a machine-readable representation of the arXiv.org collection of preprint articles. We explore fifty author-annotated categories and empirically motivate a task design of grouping 10.5 million annotated paragraphs into thirteen classes. We demonstrate that the task setup aligns with known success rates from the state of the art, peaking at a 0.91 F1-score via a BiLSTM encoder-decoder model. Additionally, we introduce a lexeme serialization for mathematical formulas, and observe that context-aware models could improve when also trained on the symbolic modality. Finally, we discuss the limitations of both data and task design, and outline potential directions towards increasingly complex models of scientific discourse, beyond isolated statements.
Author profiling models predict demographic characteristics of a target author based on the text that they have written. Systems of this kind will often follow a single-domain approach, in which the model is trained from a corpus of labelled texts in a given domain, and it is subsequently validated against a test corpus built from precisely the same domain. Although single-domain settings are arguably ideal, this strategy gives rise to the question of how to proceed when no suitable training corpus (i.e., a corpus that matches the test domain) is available. To shed light on this issue, this paper discusses a cross-domain gender classification task based on four domains (Facebook, crowd sourced opinions, Blogs and E-gov requests) in the Brazilian Portuguese language. A number of simple gender classification models using word- and psycholinguistics-based features alike are introduced, and their results are compared in two kinds of cross-domain setting: first, by making use of a single text source as training data for each task, and subsequently by combining multiple sources. Results confirm previous findings related to the effects of corpus size and domain similarity in English, and pave the way for further studies in the field.
We present LEDGAR, a multilabel corpus of legal provisions in contracts. The corpus was crawled and scraped from the public domain (SEC filings) and is, to the best of our knowledge, the first freely available corpus of its kind. Since the corpus was constructed semi-automatically, we apply and discuss various approaches to noise removal. Due to the rather large labelset of over 12’000 labels annotated in almost 100’000 provisions in over 60’000 contracts, we believe the corpus to be of interest for research in the field of Legal NLP, (large-scale or extreme) text classification, as well as for legal studies. We discuss several methods to sample subcopora from the corpus and implement and evaluate different automatic classification approaches. Finally, we perform transfer experiments to evaluate how well the classifiers perform on contracts stemming from outside the corpus.
Near-duplicate documents are particularly common in news media corpora. Editors often update wirefeed articles to address space constraints in print editions or to add local context; journalists often lightly modify previous articles with new information or minor corrections. Near-duplicate documents have potentially significant costs, including bloating corpora with redundant information (biasing techniques built upon such corpora) and requiring additional human and computational analytic resources for marginal benefit. Filtering near-duplicates out of a collection is thus important, and is particularly challenging in applications that require them to be filtered out in real-time with high precision. Previous near-duplicate detection methods typically work offline to identify all near-duplicate pairs in a set of documents. We propose an online system which flags a near-duplicate document by finding its most likely original. This system adapts the shingling algorithm proposed by Broder (1997), and we test it on a challenging dataset of web-based news articles. Our online system presents state-of-the-art F1-scores, and can be tuned to trade precision for recall and vice-versa. Given its performance and online nature, our method can be used in many real-world applications. We present one such application, filtering near-duplicates to improve productivity of human analysts in a situational awareness tool.
In this study, we created an automated essay scoring (AES) system for nonnative Japanese learners using an essay dataset with annotations for a holistic score and multiple trait scores, including content, organization, and language scores. In particular, we developed AES systems using two different approaches: a feature-based approach and a neural-network-based approach. In the former approach, we used Japanese-specific linguistic features, including character-type features such as “kanji” and “hiragana.” In the latter approach, we used two models: a long short-term memory (LSTM) model (Hochreiter and Schmidhuber, 1997) and a bidirectional encoder representations from transformers (BERT) model (Devlin et al., 2019), which achieved the highest accuracy in various natural language processing tasks in 2018. Overall, the BERT model achieved the best root mean squared error and quadratic weighted kappa scores. In addition, we analyzed the robustness of the outputs of the BERT model. We have released and shared this system to facilitate further research on AES for Japanese as a second language learners.
In this work we present a corpus for the evaluation of sensitive information detection approaches that addresses the need for real world sensitive information for empirical studies. Our sentence corpus contains different notions of complex sensitive information that correspond to different aspects of concern in a current trial of the Monsanto company. This paper describes the annotations process, where we both employ human annotators and furthermore create automatically inferred labels regarding technical, legal and informal communication within and with employees of Monsanto, drawing on a classification of documents by lawyers involved in the Monsanto court case. We release corpus of high quality sentences and parse trees with these two types of labels on sentence level. We characterize the sensitive information via several representative sensitive information detection models, in particular both keyword-based (n-gram) approaches and recent deep learning models, namely, recurrent neural networks (LSTM) and recursive neural networks (RecNN). Data and code are made publicly available.
Unbiased and fair reporting is an integral part of ethical journalism. Yet, political propaganda and one-sided views can be found in the news and can cause distrust in media. Both accidental and deliberate political bias affect the readers and shape their views. We contribute to a trustworthy media ecosystem by automatically identifying politically biased news articles. We introduce novel corpora annotated by two communities, i.e., domain experts and crowd workers, and we also consider automatic article labels inferred by the newspapers’ ideologies. Our goal is to compare domain experts to crowd workers and also to prove that media bias can be detected automatically. We classify news articles with a neural network and we also improve our performance in a self-supervised manner.
Having in mind the lack of work on the automatic recognition of verbal humour in Portuguese, a topic connected with fluency in a natural language, we describe the creation of three corpora, covering two styles of humour and four sources of non-humorous text, that may be used for related studies. We then report on some experiments where the created corpora were used for training and testing computational models that exploit content and linguistic features for humour recognition. The obtained results helped us taking some conclusions about this challenge and may be seen as baselines for those willing to tackle it in the future, using the same corpora.
Fact-checking information before publication has long been a core task for journalists, but recent times have seen the emergence of dedicated news items specifically aimed at fact-checking after publication. This relatively new form of fact-checking receives a fair amount of attention from academics, with current research focusing mostly on journalists’ motivations for publishing post-hoc fact-checks, the effects of fact-checking on the perceived accuracy of false claims, and the creation of computational tools for automatic fact-checking. In this paper, we propose to study fact-checks from a corpus linguistic perspective. This will enable us to gain insight in the scope and contents of fact-checks, to investigate what fact-checks can teach us about the way in which science appears (incorrectly) in the news, and to see how fact-checks behave in the science communication landscape. We report on the creation of FactCorp, a 1,16 million-word corpus containing 1,974 fact-checks from three major Dutch newspapers. We also present results of several exploratory analyses, including a rhetorical moves analysis, a qualitative content elements analysis, and keyword analyses. Through these analyses, we aim to demonstrate the wealth of possible applications that FactCorp allows, thereby stressing the importance of creating such resources.
Independently of the medial representation (written/spoken), language can exhibit characteristics of conceptual orality or literacy, which mainly manifest themselves on the lexical or syntactic level. In this paper we aim at automatically identifying conceptually-oral historical texts, with the ultimate goal of gaining knowledge about spoken data of historical time stages. We apply a set of general linguistic features that have been proven to be effective for the classification of modern language data to historical German texts from various registers. Many of the features turn out to be equally useful in determining the conceptuality of historical data as they are for modern data, especially the frequency of different types of pronouns and the ratio of verbs to nouns. Other features like sentence length, particles or interjections point to peculiarities of the historical data and reveal problems with the adoption of a feature set that was developed on modern language data.
Identifying statements related to suicidal behaviour in psychiatric electronic health records (EHRs) is an important step when modeling that behaviour, and when assessing suicide risk. We apply a deep neural network based classification model with a lightweight context encoder, to classify sentence level suicidal behaviour in EHRs. We show that incorporating information from sentences to left and right of the target sentence significantly improves classification accuracy. Our approach achieved the best performance when classifying suicidal behaviour in Autism Spectrum Disorder patient records. The results could have implications for suicidality research and clinical surveillance.
Film age appropriateness classification is an important problem with a significant societal impact that has so far been out of the interest of Natural Language Processing and Machine Learning researchers. To this end, we have collected a corpus of 17000 films along with their age ratings. We use the textual contents in an experiment to predict the correct age classification for the United States (G, PG, PG-13, R and NC-17) and the United Kingdom (U, PG, 12A, 15, 18 and R18). Our experiments indicate that gradient boosting machines beat FastText and various Deep Learning architectures. We reach an overall accuracy of 79.3% for the US ratings compared to a projected super human accuracy of 84%. For the UK ratings, we reach an overall accuracy of 65.3% (UK) compared to a projected super human accuracy of 80.0%.
Movies help us learn and inspire societal change. But they can also contain objectionable content that negatively affects viewers’ behaviour, especially children. In this paper, our goal is to predict the suitability of movie content for children and young adults based on scripts. The criterion that we use to measure suitability is the MPAA rating that is specifically designed for this purpose. We create a corpus for movie MPAA ratings and propose an RNN based architecture with attention that jointly models the genre and the emotions in the script to predict the MPAA rating. We achieve 81% weighted F1-score for the classification model that outperforms the traditional machine learning method by 7%.
Existing methods for different document classification tasks in the context of social networks typically only capture the semantics of texts, while ignoring the users who exchange the text and the network they form. However, some work has shown that incorporating the social network information in addition to information from language is effective for various NLP applications including sentiment analysis, inferring user attributes, and predicting inter-personal relations. In this paper, we present an empirical study of email classification into “Business” and “Personal” categories. We represent the email communication using various graph structures. As features, we use both the textual information from the email content and social network information from the communication graphs. We also model the thread structure for emails. We focus on detecting personal emails, and we evaluate our methods on two corpora, only one of which we train on. The experimental results reveal that incorporating social network information improves over the performance of an approach based on textual information only. The results also show that considering the thread structure of emails improves the performance further. Furthermore, our approach improves over a state-of-the-art baseline which uses node embeddings based on both lexical and social network information.
This work developed a Chinese humor corpus containing 3,365 jokes collected from over 40 sources. Each joke was labeled with five levels of funniness, eight skill sets of humor, and six dimensions of intent by only one annotator. To validate the manual labels, we trained SVM (Support Vector Machine) and BERT (Bidirectional Encoder Representations from Transformers) with half of the corpus (labeled by one annotator) to predict the skill and intent labels of the other half (labeled by the other annotator). Based on two assumptions that a valid manually labeled corpus should follow, our results showed the validity for the skill and intent labels. As to the funniness label, the validation results showed that the correlation between the corpus label and user feedback rating is marginal, which implies that the funniness level is a harder annotation problem to be solved. The contribution of this work is two folds: 1) a Chinese humor corpus is developed with labels of humor skills, intents, and funniness, which allows machines to learn more intricate humor framing, effect, and amusing level to predict and respond in proper context (https://github.com/SamTseng/Chinese_Humor_MultiLabeled). 2) An approach to verify whether a minimum human labeled corpus is valid or not, which facilitates the validation of low-resource corpora.
In this paper, we present a new parallel corpus addressed to researchers, teachers, and speech therapists interested in text simplification as a means of alleviating difficulties in children learning to read. The corpus is composed of excerpts drawn from 79 authentic literary (tales, stories) and scientific (documentary) texts commonly used in French schools for children aged between 7 to 9 years old. The excerpts were manually simplified at the lexical, morpho-syntactic, and discourse levels in order to propose a parallel corpus for reading tests and for the development of automatic text simplification tools. A sample of 21 poor-reading and dyslexic children with an average reading delay of 2.5 years read a portion of the corpus. The transcripts of readings errors were integrated into the corpus with the goal of identifying lexical difficulty in the target population. By means of statistical testing, we provide evidence that the manual simplifications significantly reduced reading errors, highlighting that the words targeted for simplification were not only well-chosen but also substituted with substantially easier alternatives. The entire corpus is available for consultation through a web interface and available on demand for research purposes.
A crucial step within secondary analysis of electronic health records (EHRs) is to identify the patient cohort under investigation. While EHRs contain medical billing codes that aim to represent the conditions and treatments patients may have, much of the information is only present in the patient notes. Therefore, it is critical to develop robust algorithms to infer patients’ conditions and treatments from their written notes. In this paper, we introduce a dataset for patient phenotyping, a task that is defined as the identification of whether a patient has a given medical condition (also referred to as clinical indication or phenotype) based on their patient note. Nursing Progress Notes and Discharge Summaries from the Intensive Care Unit of a large tertiary care hospital were manually annotated for the presence of several high-context phenotypes relevant to treatment and risk of re-hospitalization. This dataset contains 1102 Discharge Summaries and 1000 Nursing Progress Notes. Each Discharge Summary and Progress Note has been annotated by at least two expert human annotators (one clinical researcher and one resident physician). Annotated phenotypes include treatment non-adherence, chronic pain, advanced/metastatic cancer, as well as 10 other phenotypes. This dataset can be utilized for academic and industrial research in medicine and computer science, particularly within the field of medical natural language processing.
Stance detection aims to determine the attitude of a given text with respect to a specific topic or claim. While stance detection has been fairly well researched in the last years, most the work has been focused on English. This is mainly due to the relative lack of annotated data in other languages. The TW-10 referendum Dataset released at IberEval 2018 is a previous effort to provide multilingual stance-annotated data in Catalan and Spanish. Unfortunately, the TW-10 Catalan subset is extremely imbalanced. This paper addresses these issues by presenting a new multilingual dataset for stance detection in Twitter for the Catalan and Spanish languages, with the aim of facilitating research on stance detection in multilingual and cross-lingual settings. The dataset is annotated with stance towards one topic, namely, the ndependence of Catalonia. We also provide a semi-automatic method to annotate the dataset based on a categorization of Twitter users. We experiment on the new corpus with a number of supervised approaches, including linear classifiers and deep learning methods. Comparison of our new corpus with the with the TW-1O dataset shows both the benefits and potential of a well balanced corpus for multilingual and cross-lingual research on stance detection. Finally, we establish new state-of-the-art results on the TW-10 dataset, both for Catalan and Spanish.
A major challenge in modern neural networks is the utilization of previous knowledge for new tasks in an effective manner, otherwise known as transfer learning. Fine-tuning, the most widely used method for achieving this, suffers from catastrophic forgetting. The problem is often exacerbated in natural language processing (NLP). In this work, we assess progressive neural networks (PNNs) as an alternative to fine-tuning. The evaluation is based on common NLP tasks such as sequence labeling and text classification. By gauging PNNs across a range of architectures, datasets, and tasks, we observe improvements over the baselines throughout all experiments.
With the spread of online social networks, it is more and more difficult to monitor all the user-generated content. Automating the moderation process of the inappropriate exchange content on Internet has thus become a priority task. Methods have been proposed for this purpose, but it can be challenging to find a suitable dataset to train and develop them. This issue is especially true for approaches based on information derived from the structure and the dynamic of the conversation. In this work, we propose an original framework, based on the the Wikipedia Comment corpus, with comment-level abuse annotations of different types. The major contribution concerns the reconstruction of conversations, by comparison to existing corpora, which focus only on isolated messages (i.e. taken out of their conversational context). This large corpus of more than 380k annotated messages opens perspectives for online abuse detection and especially for context-based approaches. We also propose, in addition to this corpus, a complete benchmarking platform to stimulate and fairly compare scientific works around the problem of content abuse detection, trying to avoid the recurring problem of result replication. Finally, we apply two classification methods to our dataset to demonstrate its potential.
The rise of social media platforms makes it a valuable information source of recent events and users’ perspective towards them. Twitter has been one of the most important communication platforms in recent years. Event detection, one of the information extraction aspects, involves identifying specified types of events in the text. Detecting events from tweets can help to predict real-world events precisely. A serious challenge that faces Arabic event detection is the lack of Arabic datasets that can be exploited in detecting events. This paper will describe FloDusTA, which is a dataset of tweets that we have built for the purpose of developing an event detection system. The dataset contains tweets written in both Modern Standard Arabic and Saudi dialect. The process of building the dataset starting from tweets collection to annotation by human annotators will be present. The tweets are labeled with four labels: flood, dust storm, traffic accident, and non-event. The dataset was tested for classification and the result was strongly encouraging.
Social media networks have become a space where users are free to relate their opinions and sentiments which may lead to a large spreading of hatred or abusive messages which have to be moderated. This paper presents the first French corpus annotated for sexism detection composed of about 12,000 tweets. In a context of offensive content mediation on social media now regulated by European laws, we think that it is important to be able to detect automatically not only sexist content but also to identify if a message with a sexist content is really sexist (i.e. addressed to a woman or describing a woman or women in general) or is a story of sexism experienced by a woman. This point is the novelty of our annotation scheme. We also propose some preliminary results for sexism detection obtained with a deep learning approach. Our experiments show encouraging results.
The proliferation of fake news is a current issue that influences a number of important areas of society, such as politics, economy and health. In the Natural Language Processing area, recent initiatives tried to detect fake news in different ways, ranging from language-based approaches to content-based verification. In such approaches, the choice of the features for the classification of fake and true news is one of the most important parts of the process. This paper presents a study on the impact of readability features to detect fake news for the Brazilian Portuguese language. The results show that such features are relevant to the task (achieving, alone, up to 92% classification accuracy) and may improve previous classification results.
According to psycholinguistic studies, the complexity of concepts used in a text and the relations between mentioned concepts play the most important role in text understanding and maintaining reader’s interest. However, the classical approaches to automatic assessment of text complexity, and their commercial applications, take into consideration mainly syntactic and lexical complexity. Recently, we introduced the task of automatic assessment of conceptual text complexity, proposing a set of graph-based deep semantic features using DBpedia as a proxy to human knowledge. Given that such graphs can be noisy, incomplete, and computationally expensive to deal with, in this paper, we propose the use of textual features and shallow semantic features that only require entity linking. We compare the results obtained with new features with those of the state-of-the-art deep semantic features on two tasks: (1) pairwise comparison of two versions of the same text; and (2) five-level classification of texts. We find that the shallow features achieve state-of-the-art results on both tasks, significantly outperforming performances of the deep semantic features on the five-level classification task. Interestingly, the combination of the shallow and deep semantic features lead to a significant improvement of the performances on that task.
In recent years, the increasing interest in the development of automatic approaches for unmasking deception in online sources led to promising results. Nonetheless, among the others, two major issues remain still unsolved: the stability of classifiers performances across different domains and languages. Tackling these issues is challenging since labelled corpora involving multiple domains and compiled in more than one language are few in the scientific literature. For filling this gap, in this paper we introduce DecOp (Deceptive Opinions), a new language resource developed for automatic deception detection in cross-domain and cross-language scenarios. DecOp is composed of 5000 examples of both truthful and deceitful first-person opinions balanced both across five different domains and two languages and, to the best of our knowledge, is the largest corpus allowing cross-domain and cross-language comparisons in deceit detection tasks. In this paper, we describe the collection procedure of the DecOp corpus and his main characteristics. Moreover, the human performance on the DecOp test-set and preliminary experiments by means of machine learning models based on Transformer architecture are shown.
The understanding of a text by a reader or listener is conditioned by the adequacy of the text’s characteristics with the person’s capacities and knowledge. This adequacy is critical in the case of a child since her/his cognitive and linguistic skills are still under development. Hence, in this paper, we present and study an original natural language processing (NLP) task which consists in predicting the age from which a text can be understood by someone. To do so, this paper first exhibits features derived from the psycholinguistic domain, as well as some coming from related NLP tasks. Then, we propose a set of neural network models and compare them on a dataset of French texts dedicated to young or adult audiences. To circumvent the lack of data, we study the idea to predict ages at the sentence level. The experiments first show that the sentence-based age recommendations can be efficiently merged to predict text-based recommendations. Then, we also demonstrate that the age predictions returned by our best model are better than those provided by psycholinguists. Finally, the paper investigates the impact of the various features used in these results.
Existing research on fairness evaluation of document classification models mainly uses synthetic monolingual data without ground truth for author demographic attributes. In this work, we assemble and publish a multilingual Twitter corpus for the task of hate speech detection with inferred four author demographic factors: age, country, gender and race/ethnicity. The corpus covers five languages: English, Italian, Polish, Portuguese and Spanish. We evaluate the inferred demographic labels with a crowdsourcing platform, Figure Eight. To examine factors that can cause biases, we take an empirical analysis of demographic predictability on the English corpus. We measure the performance of four popular document classifiers and evaluate the fairness and bias of the baseline classifiers on the author-level demographic attributes.
This paper describes VICTOR, a novel dataset built from Brazil’s Supreme Court digitalized legal documents, composed of more than 45 thousand appeals, which includes roughly 692 thousand documents—about 4.6 million pages. The dataset contains labeled text data and supports two types of tasks: document type classification; and theme assignment, a multilabel problem. We present baseline results using bag-of-words models, convolutional neural networks, recurrent neural networks and boosting algorithms. We also experiment using linear-chain Conditional Random Fields to leverage the sequential nature of the lawsuits, which we find to lead to improvements on document type classification. Finally we compare a theme classification approach where we use domain knowledge to filter out the less informative document pages to the default one where we use all pages. Contrary to the Court experts’ expectations, we find that using all available data is the better method. We make the dataset available in three versions of different sizes and contents to encourage explorations of better models and techniques.
The Web archived data usually contains high-quality documents that are very useful for creating specialized collections of documents. To create such collections, there is a substantial need for automatic approaches that can distinguish the documents of interest for a collection out of the large collections (of millions in size) from Web Archiving institutions. However, the patterns of the documents of interest can differ substantially from one document to another, which makes the automatic classification task very challenging. In this paper, we explore dynamic fusion models to find, on the fly, the model or combination of models that performs best on a variety of document types. Our experimental results show that the approach that fuses different models outperforms individual models and other ensemble methods on three datasets.
For better understanding how people write texts, it is fundamental to examine how a particular aspect (e.g., subjectivity, sentiment, argumentation) is exploited in a text. Analysing such an aspect of a text as a whole (i.e., through a summarised single feature) can lead to significant information loss. In this paper, we propose a novel method of representing and analysing texts that consider how an aspect behaves throughout the text. We represent the texts by aspect flows for capturing all the aspect behaviour. Then, inspired by the resemblance between these flows format and a sound waveform, we fragment them into frames and calculate an adaptation of audio analysis features, named here Audio-Like Features, as a way of analysing the texts. The results of the conducted classification tasks reveal that our approach can surpass methods based on summarised features. We also show that a detailed examination of the Audio-Like Features can lead to a more profound knowledge about the represented texts.
The spread of biased news and its consumption by the readers has become a considerable issue. Researchers from multiple domains including social science and media studies have made efforts to mitigate this media bias issue. Specifically, various techniques ranging from natural language processing to machine learning have been used to help determine news bias automatically. However, due to the lack of publicly available datasets in this field, especially ones containing labels concerning bias on a fine-grained level (e.g., on sentence level), it is still challenging to develop methods for effectively identifying bias embedded in new articles. In this paper, we propose a novel news bias dataset which facilitates the development and evaluation of approaches for detecting subtle bias in news articles and for understanding the characteristics of biased sentences. Our dataset consists of 966 sentences from 46 English-language news articles covering 4 different events and contains labels concerning bias on the sentence level. For scalability reasons, the labels were obtained based on crowd-sourcing. Our dataset can be used for analyzing news bias, as well as for developing and evaluating methods for news bias detection. It can also serve as resource for related researches including ones focusing on fake news detection.
With the tremendous success of deep learning models on computer vision tasks, there are various emerging works on the Natural Language Processing (NLP) task of Text Classification using parametric models. However, it constrains the expressability limit of the function and demands enormous empirical efforts to come up with a robust model architecture. Also, the huge parameters involved in the model causes over-fitting when dealing with small datasets. Deep Gaussian Processes (DGP) offer a Bayesian non-parametric modelling framework with strong function compositionality, and helps in overcoming these limitations. In this paper, we propose DGP models for the task of Text Classification and an empirical comparison of the performance of shallow and Deep Gaussian Process models is made. Extensive experimentation is performed on the benchmark Text Classification datasets such as TREC (Text REtrieval Conference), SST (Stanford Sentiment Treebank), MR (Movie Reviews), R8 (Reuters-8), which demonstrate the effectiveness of DGP models.
In recent years emotion detection in text has become more popular due to its potential applications in fields such as psychology, marketing, political science, and artificial intelligence, among others. While opinion mining is a well-established task with many standard data sets and well-defined methodologies, emotion mining has received less attention due to its complexity. In particular, the annotated gold standard resources available are not enough. In order to address this shortage, we present a multilingual emotion data set based on different events that took place in April 2019. We collected tweets from the Twitter platform. Then one of seven emotions, six Ekman’s basic emotions plus the “neutral or other emotions”, was labeled on each tweet by 3 Amazon MTurkers. A total of 8,409 in Spanish and 7,303 in English were labeled. In addition, each tweet was also labeled as offensive or no offensive. We report some linguistic statistics about the data set in order to observe the difference between English and Spanish speakers when they express emotions related to the same events. Moreover, in order to validate the effectiveness of the data set, we also propose a machine learning approach for automatically detecting emotions in tweets for both languages, English and Spanish.
Endowing automated agents with the ability to provide support, entertainment and interaction with human beings requires sensing of the users’ affective state. These affective states are impacted by a combination of emotion inducers, current psychological state, and various conversational factors. Although emotion classification in both singular and dyadic settings is an established area, the effects of these additional factors on the production and perception of emotion is understudied. This paper presents a new dataset, Multimodal Stressed Emotion (MuSE), to study the multimodal interplay between the presence of stress and expressions of affect. We describe the data collection protocol, the possible areas of use, and the annotations for the emotional content of the recordings. The paper also presents several baselines to measure the performance of multimodal features for emotion and stress classification.
People convey sentiments and emotions through language. To understand these affectual states is an essential step towards understanding natural language. In this paper, we propose a transfer-learning based approach to inferring the affectual state of a person from their tweets. As opposed to the traditional machine learning models which require considerable effort in designing task specific features, our model can be well adapted to the proposed tasks with a very limited amount of fine-tuning, which significantly reduces the manual effort in feature engineering. We aim to show that by leveraging the pre-learned knowledge, transfer learning models can achieve competitive results in the affectual content analysis of tweets, compared to the traditional models. As shown by the experiments on SemEval-2018 Task 1: Affect in Tweets, our model ranking 2nd, 4th and 6th place in four of its subtasks proves the effectiveness of our idea.
We are interested in the problem of understanding personal narratives (PN) - spoken or written - recollections of facts, events, and thoughts. For PNs, we define emotion carriers as the speech or text segments that best explain the emotional state of the narrator. Such segments may span from single to multiple words, containing for example verb or noun phrases. Advanced automatic understanding of PNs requires not only the prediction of the narrator’s emotional state but also to identify which events (e.g. the loss of a relative or the visit of grandpa) or people (e.g. the old group of high school mates) carry the emotion manifested during the personal recollection. This work proposes and evaluates an annotation model for identifying emotion carriers in spoken personal narratives. Compared to other text genres such as news and microblogs, spoken PNs are particularly challenging because a narrative is usually unstructured, involving multiple sub-events and characters as well as thoughts and associated emotions perceived by the narrator. In this work, we experiment with annotating emotion carriers in speech transcriptions from the Ulm State-of-Mind in Speech (USoMS) corpus, a dataset of PNs in German. We believe this resource could be used for experiments in the automatic extraction of emotion carriers from PN, a task that could provide further advancements in narrative understanding.
Manual annotation of speech corpora is expensive in both human resources and time. Furthermore, recognizing affects in spontaneous, non acted speech presents a challenge for humans and machines. The aim of the present study is to automatize the labeling of hesitant speech as a marker of expressed uncertainty. That is why, the NCCFr-corpus was manually annotated for ‘degree of hesitation’ on a continuous scale between -3 and 3 and the affective dimensions ‘activation, valence and control’. In total, 5834 chunks of the NCCFr-corpus were manually annotated. Acoustic analyses were carried out based on these annotations. Furthermore, regression models were trained in order to allow automatic prediction of hesitation for speech chunks that do not have a manual annotation. Preliminary results show that the number of filled pauses as well as vowel duration increase with the degree of hesitation, and that automatic prediction of the hesitation degree reaches encouraging RMSE results of 1.6.
Abusive texts are reaching the interests of the scientific and social community. How to automatically detect them is onequestion that is gaining interest in the natural language processing community. The main contribution of this paper is toevaluate the quality of the recently developed ”Spanish Database for cyberbullying prevention” for the purpose of trainingclassifiers on detecting abusive short texts. We compare classical machine learning techniques to the use of a more ad-vanced model: the contextual word embeddings in the particular case of classification of abusive short-texts for the Spanishlanguage. As contextual word embeddings, we use Bidirectional Encoder Representation from Transformers (BERT), pro-posed at the end of 2018. We show that BERT mostly outperforms classical techniques. Far beyond the experimentalimpact of our research, this project aims at planting the seeds for an innovative technological tool with a high potentialsocial impact and aiming at being part of the initiatives in artificial intelligence for social good.
A fundamental essence for emotional speech analysis towards emotion recognition is a good database. Database collected from natural scenarios consists of spontaneous emotions, but there are several issues in collection of such database. Other than the privacy and legal related concerns, there is no control over environment at the background. As it is difficult to collect data from natural scenarios, many research groups have collected data from semi-natural or designed procedures. In this paper, a new emotional speech database named IIIT-H TEMD (International Institute of Information Technology-Hyderabad Telugu Emotional Database) is collected using designed drama situations from actors and non-actors. Utterances are manually annotated using a hybrid strategy by giving the context to one of the listeners. As some of the data collection studies in the literature recommend for actors, analysis of actors data versus non-actors data is carried out for their significance. The total size of the dataset is about 5 hours, which makes it an useful resource for the emotional speech analysis.
One of the main challenges in the field of Embodied Conversational Agent (ECA) is to generate socially believable agents. The common strategy for agent behaviour synthesis is to rely on dedicated corpus analysis. Such a corpus is composed of multimedia files of socio-emotional behaviors which have been annotated by external observers. The underlying idea is to identify interaction information for the agent’s socio-emotional behavior by checking whether the intended socio-emotional behavior is actually perceived by humans. Then, the annotations can be used as learning classes for machine learning algorithms applied to the social signals. This paper introduces the POTUS Corpus composed of high-quality audio-video files of political addresses to the American people. Two protagonists are present in this database. First, it includes speeches of former president Barack Obama to the American people. Secondly, it provides videos of these same speeches given by a virtual agent named Rodrigue. The ECA reproduces the original address as closely as possible using social signals automatically extracted from the original one. Both are annotated for social attitudes, providing information about the stance observed in each file. It also provides the social signals automatically extracted from Obama’s addresses used to generate Rodrigue’s ones.
Most research on emotion analysis from text focuses on the task of emotion classification or emotion intensity regression. Fewer works address emotions as a phenomenon to be tackled with structured learning, which can be explained by the lack of relevant datasets. We fill this gap by releasing a dataset of 5000 English news headlines annotated via crowdsourcing with their associated emotions, the corresponding emotion experiencers and textual cues, related emotion causes and targets, as well as the reader’s perception of the emotion of the headline. This annotation task is comparably challenging, given the large number of classes and roles to be identified. We therefore propose a multiphase annotation procedure in which we first find relevant instances with emotional content and then annotate the more fine-grained aspects. Finally, we develop a baseline for the task of automatic prediction of semantic role structures and discuss the results. The corpus we release enables further research on emotion classification, emotion intensity prediction, emotion cause detection, and supports further qualitative studies.
The state of being alone can have a substantial impact on our lives, though experiences with time alone diverge significantly among individuals. Psychologists distinguish between the concept of solitude, a positive state of voluntary aloneness, and the concept of loneliness, a negative state of dissatisfaction with the quality of one’s social interactions. Here, for the first time, we conduct a large-scale computational analysis to explore how the terms associated with the state of being alone are used in online language. We present SOLO (State of Being Alone), a corpus of over 4 million tweets collected with query terms solitude, lonely, and loneliness. We use SOLO to analyze the language and emotions associated with the state of being alone. We show that the term solitude tends to co-occur with more positive, high-dominance words (e.g., enjoy, bliss) while the terms lonely and loneliness frequently co-occur with negative, low-dominance words (e.g., scared, depressed), which confirms the conceptual distinctions made in psychology. We also show that women are more likely to report on negative feelings of being lonely as compared to men, and there are more teenagers among the tweeters that use the word lonely than among the tweeters that use the word solitude.
Child language studies are crucial in improving our understanding of child well-being; especially in determining the factors that impact happiness, the sources of anxiety, techniques of emotion regulation, and the mechanisms to cope with stress. However, much of this research is stymied by the lack of availability of large child-written texts. We present a new corpus of child-written text, PoKi, which includes about 62 thousand poems written by children from grades 1 to 12. PoKi is especially useful in studying child language because it comes with information about the age of the child authors (their grade). We analyze the words in PoKi along several emotion dimensions (valence, arousal, dominance) and discrete emotions (anger, fear, sadness, joy). We use non-parametric regressions to model developmental differences from early childhood to late-adolescence. Results show decreases in valence that are especially pronounced during mid-adolescence, while arousal and dominance peaked during adolescence. Gender differences in the developmental trajectory of emotions are also observed. Our results support and extend the current state of emotion development research.
We present a new corpus, named AlloSat, composed of real-life call center conversations in French that is continuously annotated in frustration and satisfaction. This corpus has been set up to develop new systems able to model the continuous aspect of semantic and paralinguistic information at the conversation level. The present work focuses on the paralinguistic level, more precisely on the expression of emotions. In the call center industry, the conversation usually aims at solving the caller’s request. As far as we know, most emotional databases contain static annotations in discrete categories or in dimensions such as activation or valence. We hypothesize that these dimensions are not task-related enough. Moreover, static annotations do not enable to explore the temporal evolution of emotional states. To solve this issue, we propose a corpus with a rich annotation scheme enabling a real-time investigation of the axis frustration / satisfaction. AlloSat regroups 303 conversations with a total of approximately 37 hours of audio, all recorded in real-life environments collected by Allo-Media (an intelligent call tracking company). First regression experiments, with audio features, show that the evolution of frustration / satisfaction axis can be retrieved automatically at the conversation level.
It is hard to evaluate the quality of the generated text by a generative dialogue system. Currently, dialogue evaluation relies on human judges to label the quality of the generated text. It is not a reusable mechanism that can give consistent evaluation for system developers. We believe that it is easier to get consistent results on comparing two generated dialogue by two systems and it is hard to give a consistent quality score on only one system at a time. In this paper, we propose a machine learning approach to reduce the effort of human evaluation by learning the human judgment on comparing two dialogue systems. Training from the human labeling result, the evaluation model learns which generative models is better in each dialog context. Thus, it can be used for system developers to compare the fine-tuned models over and over again without the human labor. In our experiment we find the agreement between the learned model and human judge is 70%. The experiment is conducted on comparing two attention based GRU-RNN generative models.
Detecting emotions from texts is considerably important in an NLP task, but it has the limitation of the scarcity of manually labeled data. To overcome this limitation, many researchers have annotated unlabeled data with certain frequently used annotation procedures. However, most of these studies are focused mainly on English and do not consider the characteristics of the Korean language. In this paper, we present a Korean-specific annotation procedure, which consists of two parts, namely n-gram-based distant supervision and Korean-specific-feature-based distant supervision. We leverage the distant supervision with the n-gram and Korean emotion lexicons. Then, we consider the Korean-specific emotion features. Through experiments, we showed the effectiveness of our procedure by comparing with the KTEA dataset. Additionally, we constructed a large-scale emotion-labeled dataset, Korean Movie Review Emotion (KMRE) Dataset, using our procedure. In order to construct our dataset, we used a large-scale sentiment movie review corpus as the unlabeled dataset. Moreover, we used a Korean emotion lexicon provided by KTEA. We also performed an emotion classification task and a human evaluation on the KMRE dataset.
In the case of using a deep learning (machine learning) framework for emotion classification, one significant difficulty faced is the requirement of building a large, emotion corpus in which each sentence is assigned emotion labels. As a result, there is a high cost in terms of time and money associated with the construction of such a corpus. Therefore, this paper proposes a method of creating a semi-automatically constructed emotion corpus. For the purpose of this study sentences were mined from Twitter using some emotional seed words that were selected from a dictionary in which the emotion words were well-defined. Tweets were retrieved by one emotional seed word, and the retrieved sentences were assigned emotion labels based on the emotion category of the seed word. It was evident from the findings that the deep learning-based emotion classification model could not achieve high levels of accuracy in emotion classification because the semi-automatically constructed corpus had many errors when assigning emotion labels. In this paper, therefore, an approach for improving the quality of the emotion labels by automatically correcting the errors of emotion labels is proposed and tested. The experimental results showed that the proposed method worked well, and the classification accuracy rate was improved to 55.1% from 44.9% on the Twitter emotion classification task.
A suicide note is usually written shortly before the suicide and it provides a chance to comprehend the self-destructive state of mind of the deceased. From a psychological point of view, suicide notes have been utilized for recognizing the motive behind the suicide. To the best of our knowledge, there is no openly accessible suicide note corpus at present, making it challenging for the researchers and developers to deep dive into the area of mental health assessment and suicide prevention. In this paper, we create a fine-grained emotion annotated corpus (CEASE) of suicide notes in English and develop various deep learning models to perform emotion detection on the curated dataset. The corpus consists of 2393 sentences from around 205 suicide notes collected from various sources. Each sentence is annotated with a particular emotion class from a set of 15 fine-grained emotion labels, namely (forgiveness, happiness_peacefulness, love, pride, hopefulness, thankfulness, blame, anger, fear, abuse, sorrow, hopelessness, guilt, information, instructions). For the evaluation, we develop an ensemble architecture, where the base models correspond to three supervised deep learning models, namely Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM). We obtain the highest test accuracy of 60.17% and cross-validation accuracy of 60.32%
This paper describes the training of a general-purpose German sentiment classification model. Sentiment classification is an important aspect of general text analytics. Furthermore, it plays a vital role in dialogue systems and voice interfaces that depend on the ability of the system to pick up and understand emotional signals from user utterances. The presented study outlines how we have collected a new German sentiment corpus and then combined this corpus with existing resources to train a broad-coverage German sentiment model. The resulting data set contains 5.4 million labelled samples. We have used the data to train both, a simple convolutional and a transformer-based classification model and compared the results achieved on various training configurations. The model and the data set will be published along with this paper.
The classification of implicit emotions in text has always been a great challenge to emotion processing. Even though the majority of emotion expressed implicitly, most previous attempts at emotions have focused on the examination of explicit emotions. The poor performance of existing emotion identification and classification models can partly be attributed to the disregard of implicit emotions. In view of this, this paper presents the development of a Chinese event-comment social media emotion corpus. The corpus deals with both explicit and implicit emotions with more emphasis being placed on the implicit ones. This paper specifically describes the data collection and annotation of the corpus. An annotation scheme has been proposed for the annotation of emotion-related information including the emotion type, the emotion cause, the emotion reaction, the use of rhetorical question, the opinion target (i.e. the semantic role in an event that triggers an emotion), etc. Corpus data shows that the annotated items are of great value to the identification of implicit emotions. We believe that the corpus will be a useful resource for both explicit and implicit emotion classification and detection as well as event classification.
Seeing the myriad of existing emotion models, with the categorical versus dimensional opposition the most important dividing line, building an emotion-annotated corpus requires some well thought-out strategies concerning framework choice. In our work on automatic emotion detection in Dutch texts, we investigate this problem by means of two case studies. We find that the labels joy, love, anger, sadness and fear are well-suited to annotate texts coming from various domains and topics, but that the connotation of the labels strongly depends on the origin of the texts. Moreover, it seems that information is lost when an emotional state is forcedly classified in a limited set of categories, indicating that a bi-representational format is desirable when creating an emotion corpus.
Most approaches to emotion analysis of social media, literature, news, and other domains focus exclusively on basic emotion categories as defined by Ekman or Plutchik. However, art (such as literature) enables engagement in a broader range of more complex and subtle emotions. These have been shown to also include mixed emotional responses. We consider emotions in poetry as they are elicited in the reader, rather than what is expressed in the text or intended by the author. Thus, we conceptualize a set of aesthetic emotions that are predictive of aesthetic appreciation in the reader, and allow the annotation of multiple labels per line to capture mixed emotions within their context. We evaluate this novel setting in an annotation experiment both with carefully trained experts and via crowdsourcing. Our annotation with experts leads to an acceptable agreement of k = .70, resulting in a consistent dataset for future large scale analysis. Finally, we conduct first emotion classification experiments based on BERT, showing that identifying aesthetic emotions is challenging in our data, with up to .52 F1-micro on the German subset. Data and resources are available at https://github.com/tnhaider/poetry-emotion.
Despite the excellent performance of black box approaches to modeling sentiment and emotion, lexica (sets of informative words and associated weights) that characterize different emotions are indispensable to the NLP community because they allow for interpretable and robust predictions. Emotion analysis of text is increasing in popularity in NLP; however, manually creating lexica for psychological constructs such as empathy has proven difficult. This paper automatically creates empathy word ratings from document-level ratings. The underlying problem of learning word ratings from higher-level supervision has to date only been addressed in an ad hoc fashion and has not used deep learning methods. We systematically compare a number of approaches to learning word ratings from higher-level supervision against a Mixed-Level Feed Forward Network (MLFFN), which we find performs best, and use the MLFFN to create the first-ever empathy lexicon. We then use Signed Spectral Clustering to gain insights into the resulting words. The empathy and distress lexica are publicly available at: http://www.wwbp.org/lexica.html.
Methods for learning sentence representations have been actively developed in recent years. However, the lack of pre-trained models and datasets annotated at the sentence level has been a problem for low-resource languages such as Polish which led to less interest in applying these methods to language-specific tasks. In this study, we introduce two new Polish datasets for evaluating sentence embeddings and provide a comprehensive evaluation of eight sentence representation methods including Polish and multilingual models. We consider classic word embedding models, recently developed contextual embeddings and multilingual sentence encoders, showing strengths and weaknesses of specific approaches. We also examine different methods of aggregating word vectors into a single sentence vector.
Disfluent speech has been previously addressed from two main perspectives: the clinical perspective focusing on diagnostic, and the Natural Language Processing (NLP) perspective aiming at modeling these events and detect them for downstream tasks. In addition, previous works often used different metrics depending on whether the input features are text or speech, making it difficult to compare the different contributions. Here, we introduce a new evaluation framework for disfluency detection inspired by the clinical and NLP perspective together with the theory of performance from (Clark, 1996) which distinguishes between primary and collateral tracks. We introduce a novel forced-aligned disfluency dataset from a corpus of semi-directed interviews, and present baseline results directly comparing the performance of text-based features (word and span information) and speech-based (acoustic-prosodic information). Finally, we introduce new audio features inspired by the word-based span features. We show experimentally that using these features outperformed the baselines for speech-based predictions on the present dataset.
Atypical speech productions, regardless of their origins (accents, learning, pathology), need to be assessed with regard to “typical” or “expected” productions. Evaluation is necessarily based on comparisons between linguistic forms produced and linguistic forms expected. In the field of speech disorders, the intelligibility of a patient is evaluated in order to measure the functional impact of his/her pathology on his/her oral communication. The usual method is to transcribe orthographic linguistic forms perceived and to assign a global and imprecise rating based on their correctness or incorrect. To obtain a more precise evaluation of the production deviations, we propose a measurement method based on phonological transcriptions. An algorithm computes automatically and finely the distances between the phonological forms produced and expected from cost matrices based on the differences of features between phonemes. A first test of this method among a large population of healthy speakers and patients treated for cancer of the oral and pharyngeal cavities has proved its validity.
In this paper, to evaluate text coherence, we propose the paragraph ordering task as well as conducting sentence ordering. We collected four distinct corpora from different domains on which we investigate the adaptation of existing sentence ordering methods to a paragraph ordering task. We also compare the learnability and robustness of existing models by artificially creating mini datasets and noisy datasets respectively and verifying the efficiency of established models under these circumstances. Furthermore, we carry out human evaluation on the rearranged passages from two competitive models and confirm that WLCS-l is a better metric performing significantly higher correlations with human rating than τ , the most prevalent metric used before. Results from these evaluations show that except for certain extreme conditions, the recurrent graph neural network-based model is an optimal choice for coherence modeling.
To assess the robustness of NER systems, we propose an evaluation method that focuses on subsets of tokens that represent specific sources of errors: unknown words and label shift or ambiguity. These subsets provide a system-agnostic basis for evaluating specific sources of NER errors and assessing room for improvement in terms of robustness. We analyze these subsets of challenging tokens in two widely-used NER benchmarks, then exploit them to evaluate NER systems in both in-domain and out-of-domain settings. Results show that these challenging tokens explain the majority of errors made by modern NER systems, although they represent only a small fraction of test tokens. They also indicate that label shift is harder to deal with than unknown words, and that there is much more room for improvement than the standard NER evaluation procedure would suggest. We hope this work will encourage NLP researchers to adopt rigorous and meaningful evaluation methods, and will help them develop more robust models.
Formulaic expressions, such as ‘in this paper we propose’, are used by authors of scholarly papers to perform communicative functions; the communicative function of the present example is ‘stating the aim of the paper’. Collecting such expressions and pairing them with their communicative functions would be highly valuable for various tasks, particularly for writing assistance. However, such collection and paring in a principled and automated manner would require high-quality annotated data, which are not available. In this study, we address this shortcoming by creating a manually annotated dataset for detecting communicative functions in sentences. Starting from a seed list of labelled formulaic expressions, we retrieved new sentences from scholarly papers in the ACL Anthology and asked multiple human evaluators to label communicative functions. To show the usefulness of our dataset, we conducted a series of experiments that determined to what extent sentence representations acquired by recent models, such as word2vec and BERT, can be employed to detect communicative functions in sentences.
The aim of evaluating children speech and language is to measure their communication skills. In particular, the speech language pathologist is interested in determining the child’s impairments in the areas of language, articulation, voice, fluency and swallowing. In literature some standardized tests have been proposed to assess and screen developmental language impairments but they require manual laborious transcription, annotation and calculation. This work is very time demanding and, also, may introduce several kinds of errors in the evaluation phase and non-uniform evaluations. In order to help therapists, a system performing automated evaluation is proposed. Providing as input the correct sentence and the sentence produced by patients, the technique evaluates the level of the verbal production and returns a score. The main phases of the method concern an ad-hoc transformation of the produced sentence in the reference sentence and in the evaluation of the cost of this transformation. Since the cost function is related to many weights, a learning phase is defined to automatically set such weights.
Text representations are critical for modern natural language processing. One form of text representation, sense-specific embeddings, reflect a word’s sense in a sentence better than single-prototype word embeddings tied to each type. However, existing sense representations are not uniformly better: although they work well for computer-centric evaluations, they fail for human-centric tasks like inspecting a language’s sense inventory. To expose this discrepancy, we propose a new coherence evaluation for sense embeddings. We also describe a minimal model (Gumbel Attention for Sense Induction) optimized for discovering interpretable sense representations that are more coherent than existing sense embeddings.
Summarizing data samples by quantitative measures has a long history, with descriptive statistics being a case in point. However, as natural language processing methods flourish, there are still insufficient characteristic metrics to describe a collection of texts in terms of the words, sentences, or paragraphs they comprise. In this work, we propose metrics of diversity, density, and homogeneity that quantitatively measure the dispersion, sparsity, and uniformity of a text collection. We conduct a series of simulations to verify that each metric holds desired properties and resonates with human intuitions. Experiments on real-world datasets demonstrate that the proposed characteristic metrics are highly correlated with text classification performance of a renowned model, BERT, which could inspire future applications.
Automatically analyzing events in a large amount of text is crucial for situation awareness and decision making. Previous approaches treat event extraction as “one size fits all” with an ontology defined a priori. The resulted extraction models are built just for extracting those types in the ontology. These approaches cannot be easily adapted to new event types nor new domains of interest. To accommodate personalized event-centric information needs, this paper introduces the few-shot Event Mention Retrieval (EMR) task: given a user-supplied query consisting of a handful of event mentions, return relevant event mentions found in a corpus. This formulation enables “query by example”, which drastically lowers the bar of specifying event-centric information needs. The retrieval setting also enables fuzzy search. We present an evaluation framework leveraging existing event datasets such as ACE. We also develop a Siamese Network approach, and show that it performs better than ad-hoc retrieval models in the few-shot EMR setting.
Automatic generation of reading comprehension questions is a topic receiving growing interest in the NLP community, but there is currently no consensus on evaluation metrics and many approaches focus on linguistic quality only while ignoring the pedagogic value and appropriateness of questions. This paper overcomes such weaknesses by a new evaluation scheme where questions from the questionnaire are structured in a hierarchical way to avoid confronting human annotators with evaluation measures that do not make sense for a certain question. We show through an annotation study that our scheme can be applied, but that expert annotators with some level of expertise are needed. We also created and evaluated two new evaluation data sets from the biology domain for Basque and German, composed of questions written by people with an educational background, which will be publicly released. Results show that manually generated questions are in general both of higher linguistic as well as pedagogic quality and that among the human generated questions, teacher-generated ones tend to be most useful.
Timeline summarization (TLS) generates a dated overview of real-world events based on event-specific corpora. The two standard datasets for this task were collected using Google searches for news reports on given events. Not only is this IR method not reproducible at different search times, it also uses components (such as document popularity) that are not always available for any large news corpus. It is unclear how TLS algorithms fare when provided with event corpora collected with varying IR methods. We therefore construct event-specific corpora from a large static background corpus, the newsroom dataset, using differing, relatively simple IR methods based on raw text alone. We show that the choice of IR method plays a crucial role in the performance of various TLS algorithms. A weak TLS algorithm can even match a stronger one by employing a stronger IR method in the data collection phase. Furthermore, the results of TLS systems are often highly sensitive to additional sentence filtering. We consequently advocate for integrating IR into the development of TLS systems and having a common static background corpus for evaluation of TLS systems.
The traditional approach of querying a relational database is via a formal language, namely SQL. Recent developments in the design of natural language interfaces to databases show promising results for querying either with keywords or with full natural language queries and thus render relational databases more accessible to non-tech savvy users. Such enhanced relational databases basically use a search paradigm which is commonly used in the field of information retrieval. However, the way systems are evaluated in the database and the information retrieval communities often differs due to a lack of common benchmarks. In this paper, we provide an adapted benchmark data set that is based on a test collection originally used to evaluate information retrieval systems. The data set contains 45 information needs developed on the Internet Movie Database (IMDb), including corresponding relevance assessments. By mapping this benchmark data set to a relational database schema, we enable a novel way of directly comparing database search techniques with information retrieval. To demonstrate the feasibility of our approach, we present an experimental evaluation that compares SODA, a keyword-enabled relational database system, against the Terrier information retrieval system and thus lays the foundation for a future discussion of evaluating database systems that support natural language interfaces.
As the demand for explainable deep learning grows in the evaluation of language technologies, the value of a principled grounding for those explanations grows as well. Here we study the state-of-the-art in explanation for neural models for NLP tasks from the viewpoint of philosophy of science. We focus on recent evaluation work that finds brittleness in explanations obtained through attention mechanisms. We harness philosophical accounts of explanation to suggest broader conclusions from these studies. From this analysis, we assert the impossibility of causal explanations from attention layers over text data. We then introduce NLP researchers to contemporary philosophy of science theories that allow robust yet non-causal reasoning in explanation, giving computer scientists a vocabulary for future research.
This paper investigates random vs. phonetically motivated reduction of linguistic material used in an intelligibility task in speech disordered populations and the subsequent impact on the discrimination classifier quantified by the area under the receiver operating characteristics curve (AUC of ROC). The comparison of obtained accuracy indexes shows that when the sample size is reduced based on a phonetic criterium—here, related to phonotactic complexity—, the classifier has a higher ranking ability than when the linguistic material is arbitrarily reduced. Crucially, downsizing the linguistic sample to about 30% of the original dataset does not diminish the discriminatory performance of the classifier. This result is of significant interest to both clinicians and patients as it validates a tool that is both reliable and efficient.
With the explosive growth in textual data, it is becoming increasingly important to summarize text automatically. Recently, generative language models have shown promise in abstractive text summarization tasks. Since these models rephrase text and thus use similar but different words as found in the summarized text, existing metrics such as ROUGE that use n-gram overlap may not be optimal. Therefore we evaluate two embedding-based evaluation metrics that are applicable to abstractive summarization: Fr ́echet embedding distance, which has been introduced recently, and angular embedding similarity, which is our proposed metric. To demonstrate the utility of both metrics, we analyze the headline generation capacity of two state-of-the-art language models: GPT-2 and ULMFiT. In particular, our proposed metric shows close relation with human judgments in our experiments and has overall better correlations with them. To provide reproducibility, the source code plus human assessments of our experiments is available on GitHub.
Recent trends in NLP research have raised an interest in linguistic code-switching (CS); modern approaches have been proposed to solve a wide range of NLP tasks on multiple language pairs. Unfortunately, these proposed methods are hardly generalizable to different code-switched languages. In addition, it is unclear whether a model architecture is applicable for a different task while still being compatible with the code-switching setting. This is mainly because of the lack of a centralized benchmark and the sparse corpora that researchers employ based on their specific needs and interests. To facilitate research in this direction, we propose a centralized benchmark for Linguistic Code-switching Evaluation (LinCE) that combines eleven corpora covering four different code-switched language pairs (i.e., Spanish-English, Nepali-English, Hindi-English, and Modern Standard Arabic-Egyptian Arabic) and four tasks (i.e., language identification, named entity recognition, part-of-speech tagging, and sentiment analysis). As part of the benchmark centralization effort, we provide an online platform where researchers can submit their results while comparing with others in real-time. In addition, we provide the scores of different popular models, including LSTM, ELMo, and multilingual BERT so that the NLP community can compare against state-of-the-art systems. LinCE is a continuous effort, and we will expand it with more low-resource languages and tasks.
In this paper, we investigate paraphrase generation in the colloquial domain. We use state-of-the-art neural machine translation models trained on the Opusparcus corpus to generate paraphrases in six languages: German, English, Finnish, French, Russian, and Swedish. We perform experiments to understand how data selection and filtering for diverse paraphrase pairs affects the generated paraphrases. We compare two different model architectures, an RNN and a Transformer model, and find that the Transformer does not generally outperform the RNN. We also conduct human evaluation on five of the six languages and compare the results to the automatic evaluation metrics BLEU and the recently proposed BERTScore. The results advance our understanding of the trade-offs between the quality and novelty of generated paraphrases, affected by the data selection method. In addition, our comparison of the evaluation methods shows that while BLEU correlates well with human judgments at the corpus level, BERTScore outperforms BLEU in both corpus and sentence-level evaluation.
Many researchers have tried to predict the accuracies of extrinsic evaluation by using intrinsic evaluation to evaluate word embedding. The relationship between intrinsic and extrinsic evaluation, however, has only been studied with simple correlation analysis, which has difficulty capturing complex cause-effect relationships and integrating external factors such as the hyperparameters of word embedding. To tackle this problem, we employ partial least squares path modeling (PLS-PM), a method of structural equation modeling developed for causal analysis. We propose a causal diagram consisting of the evaluation results on the BATS, VecEval, and SentEval datasets, with a causal hypothesis that linguistic knowledge encoded in word embedding contributes to solving downstream tasks. Our PLS-PM models are estimated with 600 word embeddings, and we prove the existence of causal relations between linguistic knowledge evaluated on BATS and the accuracies of downstream tasks evaluated on VecEval and SentEval in our PLS-PM models. Moreover, we show that the PLS-PM models are useful for analyzing the effect of hyperparameters, including the training algorithm, corpus, dimension, and context window, and for validating the effectiveness of intrinsic evaluation.
Current intelligent systems need the expensive support of machine learning experts to sustain their performance level when used on a daily basis. To reduce this cost, i.e. remaining free from any machine learning expert, it is reasonable to implement lifelong (or continuous) learning intelligent systems that will continuously adapt their model when facing changing execution conditions. In this work, the systems are allowed to refer to human domain experts who can provide the system with relevant knowledge about the task. Nowadays, the fast growth of lifelong learning systems development rises the question of their evaluation. In this article we propose a generic evaluation methodology for the specific case of lifelong learning systems. Two steps will be considered. First, the evaluation of human-assisted learning (including active and/or interactive learning) outside the context of lifelong learning. Second, the system evaluation across time, with propositions of how a lifelong learning intelligent system should be evaluated when including human assisted learning or not.
This paper examines the procedure for lexico-semantic annotation of the Basic Corpus of Polish Metaphors that is the first step for annotating metaphoric expressions occurring in it. The procedure involves correcting the morphosyntactic annotation of part of the corpus that is automatically annotated on the morphosyntactic level. The main procedure concerns annotation of adjectives, adverbs, nouns and verbs (including gerunds and participles), including abbreviations of the words that belong to the above classes. It is composed of three steps: deciding whether a particular occurrence of a word is asemantic (e.g. anaphoric or strictly grammatical), whether we are dealing with a multi-word expression, reciprocal usages of the się marker and pluralia tantum, which may involve annotation with two lexical units (having two different lemmas) for a single token. We propose an interannotator agreement statistics adequate for this procedure. Finally, we discuss the preliminary results of annotation of a fragment of the corpus.
Determining and correcting spelling and grammar errors in text is an important but surprisingly difficult task. There are several reasons why this remains challenging. Errors may consist of simple typing errors like deleted, substituted, or wrongly inserted letters, but may also consist of word confusions where a word was replaced by another one. In addition, words may be erroneously split into two parts or get concatenated. Some words can contain hyphens, because they were split at the end of a line or are compound words with a mandatory hyphen. In this paper, we provide an extensive evaluation of 14 spelling correction tools on a common benchmark. In particular, the evaluation provides a detailed comparison with respect to 12 error categories. The benchmark consists of sentences from the English Wikipedia, which were distorted using a realistic error model. Measuring the quality of an algorithm with respect to these error categories requires an alignment of the original text, the distorted text and the corrected text provided by the tool. We make our benchmark generation and evaluation tools publicly available.
This paper explores the use of Deep Learning methods for automatic estimation of quality of human translations. Automatic estimation can provide useful feedback for translation teaching, examination and quality control. Conventional methods for solving this task rely on manually engineered features and external knowledge. This paper presents an end-to-end neural model without feature engineering, incorporating a cross attention mechanism to detect which parts in sentence pairs are most relevant for assessing quality. Another contribution concerns oprediction of fine-grained scores for measuring different aspects of translation quality, such as terminological accuracy or idiomatic writing. Empirical results on a large human annotated dataset show that the neural model outperforms feature-based methods significantly. The dataset and the tools are available.
This article presents the results of the evaluation campaign of language tools available for fifteen EU-official under-resourced languages. The evaluation was conducted within the MSC ITN CLEOPATRA action that aims at building the cross-lingual event-centric knowledge processing on top of the application of linguistic processing chains (LPCs) for at least 24 EU-official languages. In this campaign, we concentrated on three existing NLP platforms (Stanford CoreNLP, NLP Cube, UDPipe) that all provide models for under-resourced languages and in this first run we covered 15 under-resourced languages for which the models were available. We present the design of the evaluation campaign and present the results as well as discuss them. We considered the difference between reported and our tested results within a single percentage point as being within the limits of acceptable tolerance and thus consider this result as reproducible. However, for a number of languages, the results are below what was reported in the literature, and in some cases, our testing results are even better than the ones reported previously. Particularly problematic was the evaluation of NERC systems. One of the reasons is the absence of universally or cross-lingually applicable named entities classification scheme that would serve the NERC task in different languages analogous to the Universal Dependency scheme in parsing task. To build such a scheme has become one of our the future research directions.
This paper presents the first ever comprehensive evaluation of different types of word embeddings for Sinhala language. Three standard word embedding models, namely, Word2Vec (both Skipgram and CBOW), FastText, and Glove are evaluated under two types of evaluation methods: intrinsic evaluation and extrinsic evaluation. Word analogy and word relatedness evaluations were performed in terms of intrinsic evaluation, while sentiment analysis and part-of-speech (POS) tagging were conducted as the extrinsic evaluation tasks. Benchmark datasets used for intrinsic evaluations were carefully crafted considering specific linguistic features of Sinhala. In general, FastText word embeddings with 300 dimensions reported the finest accuracies across all the evaluation tasks, while Glove reported the lowest results.
There has been significant progress in recent years in the field of Natural Language Processing thanks to the introduction of the Transformer architecture. Current state-of-the-art models, via a large number of parameters and pre-training on massive text corpus, have shown impressive results on several downstream tasks. Many researchers have studied previous (non-Transformer) models to understand their actual behavior under different scenarios, showing that these models are taking advantage of clues or failures of datasets and that slight perturbations on the input data can severely reduce their performance. In contrast, recent models have not been systematically tested with adversarial-examples in order to show their robustness under severe stress conditions. For that reason, this work evaluates three Transformer-based models (RoBERTa, XLNet, and BERT) in Natural Language Inference (NLI) and Question Answering (QA) tasks to know if they are more robust or if they have the same flaws as their predecessors. As a result, our experiments reveal that RoBERTa, XLNet and BERT are more robust than recurrent neural network models to stress tests for both NLI and QA tasks. Nevertheless, they are still very fragile and demonstrate various unexpected behaviors, thus revealing that there is still room for future improvement in this field.
Relation Extraction is a fundamental NLP task. In this paper we investigate the impact of underlying text representation on the performance of neural classification models in the task of Brand-Product relation extraction. We also present the methodology of preparing annotated textual corpora for this task and we provide valuable insight into the properties of Brand-Product relations existing in textual corpora. The problem is approached from a practical angle of applications Relation Extraction in facilitating commercial Internet monitoring.
We discuss methodological choices in contrastive and diagnostic evaluation in meaning representation parsing, i.e. mapping from natural language utterances to graph-based encodings of its semantic structure. Drawing inspiration from earlier work in syntactic dependency parsing, we transfer and refine several quantitative diagnosis techniques for use in the context of the 2019 shared task on Meaning Representation Parsing (MRP). As in parsing proper, moving evaluation from simple rooted trees to general graphs brings along its own range of challenges. Specifically, we seek to begin to shed light on relative strenghts and weaknesses in different broad families of parsing techniques. In addition to these theoretical reflections, we conduct a pilot experiment on a selection of top-performing MRP systems and one of the five meaning representation frameworks in the shared task. Empirical results suggest that the proposed methodology can be meaningfully applied to parsing into graph-structured target representations, uncovering hitherto unknown properties of the different systems that can inform future development and cross-fertilization across approaches.
In this paper, we design headword-oriented entity linking (HEL), a specialized entity linking problem in which only the headwords of the entities are to be linked to knowledge bases; mention scopes of the entities do not need to be identified in the problem setting. This special task is motivated by the fact that in many articles referring to specific products, the complete full product names are rarely written; instead, they are often abbreviated to shorter, irregular versions or even just to their headwords, which are usually their product types, such as “stick” or “mask” in a cosmetic context. To fully design the special task, we construct a labeled cosmetic corpus as a public benchmark for this problem, and propose a product embedding model to address the task, where each product corresponds to a dense representation to encode the different information on products and their context jointly. Besides, to increase training data, we propose a special transfer learning framework in which distant supervision with heuristic patterns is first utilized, followed by supervised learning using a small amount of manually labeled data. The experimental results show that our model provides a strong benchmark performance on the special task.
We present TableBank, a new image-based table detection and recognition dataset built with novel weak supervision from Word and Latex documents on the internet. Existing research for image-based table detection and recognition usually fine-tunes pre-trained models on out-of-domain data with a few thousand human-labeled examples, which is difficult to generalize on real-world applications. With TableBank that contains 417K high quality labeled tables, we build several strong baselines using state-of-the-art models with deep neural networks. We make TableBank publicly available and hope it will empower more deep learning approaches in the table detection and recognition task. The dataset and models can be downloaded from https://github.com/doc-analysis/TableBank.
Over the past years, deep learning methods allowed for new state-of-the-art results in ad-hoc information retrieval. However such methods usually require large amounts of annotated data to be effective. Since most standard ad-hoc information retrieval datasets publicly available for academic research (e.g. Robust04, ClueWeb09) have at most 250 annotated queries, the recent deep learning models for information retrieval perform poorly on these datasets. These models (e.g. DUET, Conv-KNRM) are trained and evaluated on data collected from commercial search engines not publicly available for academic research which is a problem for reproducibility and the advancement of research. In this paper, we propose WIKIR: an open-source toolkit to automatically build large-scale English information retrieval datasets based on Wikipedia. WIKIR is publicly available on GitHub. We also provide wikIR59k: a large-scale publicly available dataset that contains 59,252 queries and 2,617,003 (query, relevant documents) pairs.
In this paper, we propose a full pipeline of analysis of a large corpus about a century of public meeting in historical Australian news papers, from construction to visual exploration. The corpus construction method is based on image processing and OCR. We digitize and transcribe texts of the specific topic of public meeting. Experiments show that our proposed method achieves a F-score of 87.8% for corpus construction. As a result, we built a content search tool for temporal and semantic content analysis.
The synthesis process is essential for achieving computational experiment design in the field of inorganic materials chemistry. In this work, we present a novel corpus of the synthesis process for all-solid-state batteries and an automated machine reading system for extracting the synthesis processes buried in the scientific literature. We define the representation of the synthesis processes using flow graphs, and create a corpus from the experimental sections of 243 papers. The automated machine-reading system is developed by a deep learning-based sequence tagger and simple heuristic rule-based relation extractor. Our experimental results demonstrate that the sequence tagger with the optimal setting can detect the entities with a macro-averaged F1 score of 0.826, while the rule-based relation extractor can achieve high performance with a macro-averaged F1 score of 0.887.
Building predictive models for information extraction from text, such as named entity recognition or the extraction of semantic relationships between named entities in text, requires a large corpus of annotated text. Wikipedia is often used as a corpus for these tasks where the annotation is a named entity linked by a hyperlink to its article. However, editors on Wikipedia are only expected to link these mentions in order to help the reader to understand the content, but are discouraged from adding links that do not add any benefit for understanding an article. Therefore, many mentions of popular entities (such as countries or popular events in history), or previously linked articles, as well as the article’s entity itself, are not linked. In this paper, we discuss WEXEA, a Wikipedia EXhaustive Entity Annotation system, to create a text corpus based on Wikipedia with exhaustive annotations of entity mentions, i.e. linking all mentions of entities to their corresponding articles. This results in a huge potential for additional annotations that can be used for downstream NLP tasks, such as Relation Extraction. We show that our annotations are useful for creating distantly supervised datasets for this task. Furthermore, we publish all code necessary to derive a corpus from a raw Wikipedia dump, so that it can be reproduced by everyone.
Entity normalization (or entity linking) is an important subtask of information extraction that links entity mentions in text to categories or concepts in a reference vocabulary. Machine learning based normalization methods have good adaptability as long as they have enough training data per reference with a sufficient quality. Distributional representations are commonly used because of their capacity to handle different expressions with similar meanings. However, in specific technical and scientific domains, the small amount of training data and the relatively small size of specialized corpora remain major challenges. Recently, the machine learning-based CONTES method has addressed these challenges for reference vocabularies that are ontologies, as is often the case in life sciences and biomedical domains. And yet, its performance is dependent on manually annotated corpus. Furthermore, like other machine learning based methods, parametrization remains tricky. We propose a new approach to address the scarcity of training data that extends the CONTES method by corpus selection, pre-processing and weak supervision strategies, which can yield high-performance results without any manually annotated examples. We also study which hyperparameters are most influential, with sometimes different patterns compared to previous work. The results show that our approach significantly improves accuracy and outperforms previous state-of-the-art algorithms.
Due to the fast pace at which research reports in behaviour change are published, researchers, consultants and policymakers would benefit from more automatic ways to process these reports. Automatic extraction of the reports’ intervention content, population, settings and their results etc. are essential in synthesising and summarising the literature. However, to the best of our knowledge, no unique resource exists at the moment to facilitate this synthesis. In this paper, we describe the construction of a corpus of published behaviour change intervention evaluation reports aimed at smoking cessation. We also describe and release the annotation of 57 entities, that can be used as an off-the-shelf data resource for tasks such as entity recognition, etc. Both the corpus and the annotation dataset are being made available to the community.
Most of the current cross-lingual transfer learning methods for Information Extraction (IE) have been only applied to name tagging. To tackle more complex tasks such as event extraction we need to transfer graph structures (event trigger linked to multiple arguments with various roles) across languages. We develop a novel share-and-transfer framework to reach this goal with three steps: (1) Convert each sentence in any language to language-universal graph structures; in this paper we explore two approaches based on universal dependency parses and complete graphs, respectively. (2) Represent each node in the graph structure with a cross-lingual word embedding so that all sentences in multiple languages can be represented with one shared semantic space. (3) Using this common semantic space, train event extractors from English training data and apply them to languages that do not have any event annotations. Experimental results on three languages (Spanish, Russian and Ukrainian) without any annotations show this framework achieves comparable performance to a state-of-the-art supervised model trained from more than 1,500 manually annotated event mentions.
Biomedical event extraction is a crucial task in order to automatically extract information from the increasingly growing body of biomedical literature. Despite advances in the methods in recent years, most event extraction systems are still evaluated in-domain and on complete event structures only. This makes it hard to determine the performance of intermediate stages of the task, such as edge detection, across different corpora. Motivated by these limitations, we present the first cross-domain study of edge detection for biomedical event extraction. We analyze differences between five existing gold standard corpora, create a standardized benchmark corpus, and provide a strong baseline model for edge detection. Experiments show a large drop in performance when the baseline is applied on out-of-domain data, confirming the need for domain adaptation methods for the task. To encourage research efforts in this direction, we make both the data and the baseline available to the research community: https://www.cosbi.eu/cfx/9985.
Risk management is a vital activity to ensure employee safety in construction projects. Various documents provide important supporting evidence, including details of previous incidents, consequences and mitigation strategies. Potential hazards may depend on a complex set of project-specific attributes, including activities undertaken, location, equipment used, etc. However, finding evidence about previous projects with similar attributes can be problematic, since information about risks and mitigations is usually hidden within and may be dispersed across a range of different free text documents. Automatic named entity recognition (NER), which identifies mentions of concepts in free text documents, is the first stage in structuring knowledge contained within them. While developing NER methods generally relies on annotated corpora, we are not aware of any such corpus targeted at concepts relevant to construction safety. In response, we have designed a novel named entity annotation scheme and associated guidelines for this domain, which covers hazards, consequences, mitigation strategies and project attributes. Four health and safety experts used the guidelines to annotate a total of 600 sentences from accident reports; an average inter-annotator agreement rate of 0.79 F-Score shows that our work constitutes an important first step towards developing tools for detailed semantic analysis of construction safety documents.
Within this work we describe a framework for the collection and summarization of information from the Web in an entity-driven manner. The framework consists of a set of appropriate workflows and the Social Web Observatory platform, which implements those workflows, supporting them through a language analysis pipeline. The pipeline includes text collection/crawling, identification of different entities, clustering of texts into events related to entities, entity-centric sentiment analysis, but also text analytics and visualization functionalities. The latter allow the user to take advantage of the gathered information as actionable knowledge: to understand the dynamics of the public opinion for a given entity over time and across real-world events. We describe the platform and the analysis functionality and evaluate the performance of the system, by allowing human users to score how the system fares in its intended purpose of summarizing entity-centered information from different sources in the Web.
Free text fields within electronic health records (EHRs) contain valuable clinical information which is often missed when conducting research using EHR databases. One such type of information is medications which are not always available in structured fields, especially in mental health records. Most use cases that require medication information also generally require the associated temporal information (e.g. current or past) and attributes (e.g. dose, route, frequency). The purpose of this study is to develop a corpus of medication annotations in mental health records. The aim is to provide a more complete picture behind the mention of medications in the health records, by including additional contextual information around them, and to create a resource for use when developing and evaluating applications for the extraction of medications from EHR text. Thus far, an analysis of temporal information related to medications mentioned in a sample of mental health records has been conducted. The purpose of this analysis was to understand the complexity of medication mentions and their associated temporal information in the free text of EHRs, with a specific focus on the mental health domain.
The Conversational Question Answering (CoQA) task involves answering a sequence of inter-related conversational questions about a contextual paragraph. Although existing approaches employ human-written ground-truth answers for answering conversational questions at test time, in a realistic scenario, the CoQA model will not have any access to ground-truth answers for the previous questions, compelling the model to rely upon its own previously predicted answers for answering the subsequent questions. In this paper, we find that compounding errors occur when using previously predicted answers at test time, significantly lowering the performance of CoQA systems. To solve this problem, we propose a sampling strategy that dynamically selects between target answers and model predictions during training, thereby closely simulating the situation at test time. Further, we analyse the severity of this phenomena as a function of the question type, conversation length and domain type.
Entity linking, as one of the fundamental tasks in natural language processing, is crucial to knowledge fusion, knowledge base construction and update. Nevertheless, in contrast to the research on entity linking for English text, which undergoes continuous development, the Chinese counterpart is still in its infancy. One prominent issue lies in publicly available annotated datasets and evaluation benchmarks, which are lacking and deficient. In specific, existing Chinese corpora for entity linking were mainly constructed from noisy short texts, such as microblogs and news headings, where long texts were largely overlooked, which yet constitute a wider spectrum of real-life scenarios. To address the issue, in this work, we build CLEEK, a Chinese corpus of multi-domain long text for entity linking, in order to encourage advancement of entity linking in languages besides English. The corpus consists of 100 documents from diverse domains, and is publicly accessible. Moreover, we devise a measure to evaluate the difficulty of documents with respect to entity linking, which is then used to characterize the corpus. Additionally, the results of two baselines and seven state-of-the-art solutions on CLEEK are reported and compared. The empirical results validate the usefulness of CLEEK and the effectiveness of proposed difficulty measure.
There is a growing interest in creating tools to assist in clinical note generation using the audio of provider-patient encounters. Motivated by this goal and with the help of providers and medical scribes, we developed an annotation scheme to extract relevant clinical concepts. We used this annotation scheme to label a corpus of about 6k clinical encounters. This was used to train a state-of-the-art tagging model. We report ontologies, labeling results, model performances, and detailed analyses of the results. Our results show that the entities related to medications can be extracted with a relatively high accuracy of 0.90 F-score, followed by symptoms at 0.72 F-score, and conditions at 0.57 F-score. In our task, we not only identify where the symptoms are mentioned but also map them to canonical forms as they appear in the clinical notes. Of the different types of errors, in about 19-38% of the cases, we find that the model output was correct, and about 17-32% of the errors do not impact the clinical note. Taken together, the models developed in this work are more useful than the F-scores reflect, making it a promising approach for practical applications.
A contract is a legal document executed by two or more parties. It is important for these parties to precisely understand their rights and obligations that are described in the contract. However, understanding the content of a contract is sometimes difficult and costly, particularly if the contract is long and complicated. Therefore, a language-processing system that can present information concerning rights and obligations found within a given contract document would help a contracting party to make better decisions. As a step toward the development of such a language-processing system, in this paper, we describe the annotated corpus of contract documents that we built. Our corpus is annotated so that a language-processing system can recognize a party’s rights and obligations. The annotated information includes the parties involved in the contract, the rights and obligations of the parties, the conditions and the exceptions under which these rights and obligations to take effect. The corpus was built based on 46 English contracts and 25 Japanese contracts drafted by lawyers. We explain how we annotated the corpus and the statistics of the corpus. We also report the results of the experiments for recognizing rights and obligations.
Analyzing the geographic movement of humans, animals, and other phenomena is a growing field of research. This research has benefited urban planning, logistics, animal migration understanding, and much more. Typically, the movement is captured as precise geographic coordinates and time stamps with Global Positioning Systems (GPS). Although some research uses computational techniques to take advantage of implicit movement in descriptions of route directions, hiking paths, and historical exploration routes, innovation would accelerate with a large and diverse corpus. We created a corpus of sentences labeled as describing geographic movement or not and including the type of entity moving. Creating this corpus proved difficult without any comparable corpora to start with, high human labeling costs, and since movement can at times be interpreted differently. To overcome these challenges, we developed an iterative process employing hand labeling, crowd voting for confirmation, and machine learning to predict more labels. By merging advances in word embeddings with traditional machine learning models and model ensembling, prediction accuracy is at an acceptable level to produce a large silver-standard corpus despite the small gold-standard corpus training set. Our corpus will likely benefit computational processing of geography in text and spatial cognition, in addition to detection of movement.
In this study, we construct a corpus of Japanese local assembly minutes. All speeches in an assembly were transcribed into a local assembly minutes based on the local autonomy law. Therefore, the local assembly minutes form an extremely large amount of text data. Our ultimate objectives were to summarize and present the arguments in the assemblies, and to use the minutes as primary information for arguments in local politics. To achieve this, we structured all statements in assembly minutes. We focused on the structure of the discussion, i.e., the extraction of question and answer pairs. We organized the shared task “QA Lab-PoliInfo” in NTCIR 14. We conducted a “segmentation task” to identify the scope of one question and answer in the minutes as a sub task of the shared task. For the segmentation task, 24 runs from five teams were submitted. Based on the obtained results, the best recall was 1.000, best precision was 0.940, and best F-measure was 0.895.
The voice of the customer has for a long time been a key focus of businesses in all domains. It has received a lot of attention from the research community in Natural Language Processing (NLP) resulting in many approaches to analyzing customers feedback ((aspect-based) sentiment analysis, topic modeling, etc.). In the health domain, public and private bodies are increasingly prioritizing patient engagement for assessing the quality of the service given at each stage of the care. Patient and customer satisfaction analysis relate in many ways. In the domain of health particularly, a more precise and insightful analysis is needed to help practitioners locate potential issues and plan actions accordingly. We introduce here an approach to patient experience with the analysis of free text questions from the 2017 Irish National Inpatient Survey campaign using term extraction as a means to highlight important and insightful subject matters raised by patients. We evaluate the results by mapping them to a manually constructed framework following the Activity, Resource, Context (ARC) methodology (Ordenes, 2014) and specific to the health care environment, and compare our results against manual annotations done on the full 2017 dataset based on those categories.
Nature has inspired various ground-breaking technological developments in applications ranging from robotics to aerospace engineering and the manufacturing of medical devices. However, accessing the information captured in scientific biology texts is a time-consuming and hard task that requires domain-specific knowledge. Improving access for outsiders can help interdisciplinary research like Nature Inspired Engineering. This paper describes a dataset of 1,500 manually-annotated sentences that express domain-independent relations between central concepts in a scientific biology text, such as trade-offs and correlations. The arguments of these relations can be Multi Word Expressions and have been annotated with modifying phrases to form non-projective graphs. The dataset allows for training and evaluating Relation Extraction algorithms that aim for coarse-grained typing of scientific biological documents, enabling a high-level filter for engineers.
Automatic definition extraction from texts is an important task that has numerous applications in several natural language processing fields such as summarization, analysis of scientific texts, automatic taxonomy generation, ontology generation, concept identification, and question answering. For definitions that are contained within a single sentence, this problem can be viewed as a binary classification of sentences into definitions and non-definitions. Definitions in scientific literature can be generic (Wikipedia) or more formal (mathematical articles). In this paper, we focus on automatic detection of one-sentence definitions in mathematical texts, which are difficult to separate from surrounding text. We experiment with several data representations, which include sentence syntactic structure and word embeddings, and apply deep learning methods such as convolutional neural network (CNN) and recurrent neural network (RNN), in order to identify mathematical definitions. Our experiments demonstrate the superiority of CNN and its combination with RNN, applied on the syntactically-enriched input representation. We also present a new dataset for definition extraction from mathematical texts. We demonstrate that the use of this dataset for training learning models improves the quality of definition extraction when these models are then used for other definition datasets. Our experiments with different domains approve that mathematical definitions require special treatment, and that using cross-domain learning is inefficient.
Entities can be found in various text genres, ranging from tweets and web pages to user queries submitted to web search engines. Existing research either considers all entities in the text equally important, or heuristics are used to measure their salience. We believe that a key reason for the relatively limited work on entity salience is the lack of appropriate datasets. To support research on entity salience, we present a new dataset, the WikiNews Salience dataset (WN-Salience), which can be used to benchmark tasks such as entity salience detection and salient entity linking. WN-Salience is built on top of Wikinews, a Wikimedia project whose mission is to present reliable news articles. Entities in Wikinews articles are identified by the authors of the articles and are linked to Wikinews categories when they are salient or to Wikipedia pages otherwise. The dataset is built automatically, and consists of approximately 7,000 news articles, and 90,000 in-text entity annotations. We compare the WN-Salience dataset against existing datasets on the task and analyze their differences. Furthermore, we conduct experiments on entity salience detection; the results demonstrate that WN-Salience is a challenging testbed that is complementary to existing ones.
In information extraction, event extraction is one of the types that extract the specific knowledge of certain incidents from texts. Event extraction has been done on different languages text but not on one of the Semitic language, Amharic. In this study, we present a system that extracts an event from unstructured Amharic text. The system has designed by the integration of supervised machine learning and rule-based approaches. We call this system a hybrid system. The system uses the supervised machine learning to detect events from the text and the handcrafted and the rule-based rules to extract the event from the text. For the event extraction, we have been using event arguments. Event arguments identify event triggering words or phrases that clearly express the occurrence of the event. The event argument attributes can be verbs, nouns, sometimes adjectives (such as ̃rg/wedding) and time as well. The hybrid system has compared with the standalone rule-based method that is well known for event extraction. The study has shown that the hybrid system has outperformed the standalone rule-based method.
We present a comparison between deep learning and traditional machine learning methods for various NLP tasks in Italian. We carried on experiments using available datasets (e.g., from the Evalita shared tasks) on two sequence tagging tasks (i.e., named entities recognition and nominal entities recognition) and four classification tasks (i.e., lexical relations among words, semantic relations among sentences, sentiment analysis and text classification). We show that deep learning approaches outperform traditional machine learning algorithms in sequence tagging, while for classification tasks that heavily rely on semantics approaches based on feature engineering are still competitive. We think that a similar analysis could be carried out for other languages to provide an assessment of machine learning / deep learning models across different languages.
Text instructions are among the most widely used media for learning and teaching. Hence, to create assistance systems that are capable of supporting humans autonomously in new tasks, it would be immensely productive, if machines were enabled to extract task knowledge from such text instructions. In this paper, we, therefore, focus on information extraction (IE) from the instructional text in repair manuals. This brings with it the multiple challenges of information extraction from the situated and technical language in relatively long and often complex instructions. To tackle these challenges, we introduce a semi-structured dataset of repair manuals. The dataset is annotated in a large category of devices, with information that we consider most valuable for an automated repair assistant, including the required tools and the disassembled parts at each step of the repair progress. We then propose methods that can serve as baselines for this IE task: an unsupervised method based on a bags-of-n-grams similarity for extracting the needed tools in each repair step, and a deep-learning-based sequence labeling model for extracting the identity of disassembled parts. These baseline methods are integrated into a semi-automatic web-based annotator application that is also available along with the dataset.
Entities are a central element of knowledge bases and are important input to many knowledge-centric tasks including text analysis. For example, they allow us to find documents relevant to a specific entity irrespective of the underlying syntactic expression within a document. However, the entities that are commonly represented in knowledge bases are often a simplification of what is truly being referred to in text. For example, in a knowledge base, we may have an entity for Germany as a country but not for the more fuzzy concept of Germany that covers notions of German Population, German Drivers, and the German Government. Inspired by recent advances in contextual word embeddings, we introduce the concept of entity spaces - specific representations of a set of associated entities with near-identity. Thus, these entity spaces provide a handle to an amorphous grouping of entities. We developed a proof-of-concept for English showing how, through the introduction of entity spaces in the form of disambiguation pages, the recall of entity linking can be improved.
We present the WASABI Song Corpus, a large corpus of songs enriched with metadata extracted from music databases on the Web, and resulting from the processing of song lyrics and from audio analysis. More specifically, given that lyrics encode an important part of the semantics of a song, we focus here on the description of the methods we proposed to extract relevant information from the lyrics, as their structure segmentation, their topic, the explicitness of the lyrics content, the salient passages of a song and the emotions conveyed. The creation of the resource is still ongoing: so far, the corpus contains 1.73M songs with lyrics (1.41M unique lyrics) annotated at different levels with the output of the above mentioned methods. Such corpus labels and the provided methods can be exploited by music search engines and music professionals (e.g. journalists, radio presenters) to better handle large collections of lyrics, allowing an intelligent browsing, categorization and segmentation recommendation of songs.
Temporal Dependency Trees (TDTs) have emerged as an alternative to full temporal graphs for representing the temporal structure of texts, with a key advantage being that TDTs can be straightforwardly computed using adapted dependency parsers. Relative to temporal graphs, the tree form of TDTs naturally omits some fraction of temporal relationships, which intuitively should decrease the amount of temporal information available, potentially increasing temporal indeterminacy of the global ordering. We demonstrate a new method for quantifying this indeterminacy that relies on solving temporal constraint problems to extract timelines, and show that TDTs result in up to a 109% increase in temporal indeterminacy over their corresponding temporal graphs for the three corpora we examine. On average, the increase in indeterminacy is 32%, and we show that this increase is a result of the TDT representation eliminating on average only 2.4% of total temporal relations. This result suggests that small differences can have big effects in temporal graphs, and the use of TDTs must be balanced against their deficiencies, with tasks requiring an accurate global temporal ordering potentially calling for use of the full temporal graph
This paper is concerned with the goal of maintaining legal information and compliance systems: the ‘resource consumption bottleneck’ of creating semantic technologies manually. The use of automated information extraction techniques could significantly reduce this bottleneck. The research question of this paper is: How to address the resource bottleneck problem of creating specialist knowledge management systems? In particular, how to semi-automate the extraction of norms and their elements to populate legal ontologies? This paper shows that the acquisition paradox can be addressed by combining state-of-the-art general-purpose NLP modules with pre- and post-processing using rules based on domain knowledge. It describes a Semantic Role Labeling based information extraction system to extract norms from legislation and represent them as structured norms in legal ontologies. The output is intended to help make laws more accessible, understandable, and searchable in legal document management systems such as Eunomos (Boella et al., 2016).
In the paper, we focus on modeling spatial expressions in texts. We present the guidelines used to annotate the PST 2.0 (Corpus of Polish Spatial Texts) — a corpus designed for training and testing the tools for spatial expression recognition. The corpus contains a set of texts gathered from texts collected from travel blogs available under Creative Commons license. We have defined our guidelines based on three existing specifications for English (SpatialML, SpatialRole Labelling from SemEval-2013 Task 3 and ISO-Space1.4 from SpaceEval 2014). We briefly present the existing specifications and discuss what modifications have been made to adapt the guidelines to the characteristics of the Polish language. We also describe the process of data collection and manual annotation, including inter-annotator agreement calculation and corpus statistics. In the end, we present detailed statistics of the PST 2.0 corpus, which include the number of components, relations, expressions, and the most common values of spatial indicators, motion indicators, path indicators, distances, directions, and regions.
Mathematical text is written using a combination of words and mathematical expressions. This combination, along with a specific way of structuring sentences makes it challenging for state-of-art NLP tools to understand and reason on top of mathematical discourse. In this work, we propose a new NLP task, the natural premise selection, which is used to retrieve supporting definitions and supporting propositions that are useful for generating an informal mathematical proof for a particular statement. We also make available a dataset, NL-PS, which can be used to evaluate different approaches for the natural premise selection task. Using different baselines, we demonstrate the underlying interpretation challenges associated with the task.
We present Odinson, a rule-based information extraction framework, which couples a simple yet powerful pattern language that can operate over multiple representations of text, with a runtime system that operates in near real time. In the Odinson query language, a single pattern may combine regular expressions over surface tokens with regular expressions over graphs such as syntactic dependencies. To guarantee the rapid matching of these patterns, our framework indexes most of the necessary information for matching patterns, including directed graphs such as syntactic dependencies, into a custom Lucene index. Indexing minimizes the amount of expensive pattern matching that must take place at runtime. As a result, the runtime system matches a syntax-based graph traversal in 2.8 seconds in a corpus of over 134 million sentences, nearly 150,000 times faster than its predecessor.
We introduce the STEM (Science, Technology, Engineering, and Medicine) Dataset for Scientific Entity Extraction, Classification, and Resolution, version 1.0 (STEM-ECR v1.0). The STEM-ECR v1.0 dataset has been developed to provide a benchmark for the evaluation of scientific entity extraction, classification, and resolution tasks in a domain-independent fashion. It comprises abstracts in 10 STEM disciplines that were found to be the most prolific ones on a major publishing platform. We describe the creation of such a multidisciplinary corpus and highlight the obtained findings in terms of the following features: 1) a generic conceptual formalism for scientific entities in a multidisciplinary scientific context; 2) the feasibility of the domain-independent human annotation of scientific entities under such a generic formalism; 3) a performance benchmark obtainable for automatic extraction of multidisciplinary scientific entities using BERT-based neural models; 4) a delineated 3-step entity resolution procedure for human annotation of the scientific entities via encyclopedic entity linking and lexicographic word sense disambiguation; and 5) human evaluations of Babelfy returned encyclopedic links and lexicographic senses for our entities. Our findings cumulatively indicate that human annotation and automatic learning of multidisciplinary scientific concepts as well as their semantic disambiguation in a wide-ranging setting as STEM is reasonable.
Extending machine reading approaches to extract mathematical concepts and their descriptions is useful for a variety of tasks, ranging from mathematical information retrieval to increasing accessibility of scientific documents for the visually impaired. This entails segmenting mathematical formulae into identifiers and linking them to their natural language descriptions. We propose a rule-based approach for this task, which extracts LaTeX representations of formula identifiers and links them to their in-text descriptions, given only the original PDF and the location of the formula of interest. We also present a novel evaluation dataset for this task, as well as the tool used to create it.
In this paper we present a relation extraction system that given a text extracts pedagogically motivated relation types, as a previous step to obtaining a semantic representation of the text which will make possible to automatically generate questions for reading comprehension. The system maps pedagogically motivated relations with relations from ConceptNet and deploys Distant Supervision for relation extraction. We run a study on a subset of those relationships in order to analyse the viability of our approach. For that, we build a domain-specific relation extraction system and explore two relation extraction models: a state-of-the-art model based on transfer learning and a discrete feature based machine learning model. Experiments show that the neural model obtains better results in terms of F-score and we yield promising results on the subset of relations suitable for pedagogical purposes. We thus consider that distant supervision for relation extraction is a valid approach in our target domain, i.e. biology.
We present a new temporal annotation standard, THEE-TimeML, and a corpus TheeBank enabling precise temporal information extraction (TIE) for event-based surveillance (EBS) systems in the public health domain. Current EBS must estimate the occurrence time of each event based on coarse document metadata such as document publication time. Because of the complicated language and narration style of news articles, estimated case outbreak times are often inaccurate or even erroneous. Thus, it is necessary to create annotation standards and corpora to facilitate the development of TIE systems in the public health domain to address this problem.We will discuss the adaptations that have proved necessary for this domain as we present THEE-TimeML and TheeBank. Finally, we document the corpus annotation process, and demonstrate the immediate benefit to public health applications brought by the annotations.
In this paper we address the problem of providing personalised recommendations of recent scientific publications to a particular user, and explore the use of citation knowledge to do so. For this purpose, we have generated a novel dataset that captures authors’ publication history and is enriched with different forms of paper citation knowledge, namely citation graphs, citation positions, citation contexts, and citation types. Through a number of empirical experiments on such dataset, we show that the exploitation of the extracted knowledge, particularly the type of citation, is a promising approach for recommending recently published papers that may not be cited yet. The dataset, which we make publicly available, also represents a valuable resource for further investigation on academic information retrieval and filtering.
Event Extraction is an important task in the widespread field of Natural Language Processing (NLP). Though this task is adequately addressed in English with sufficient resources, we are unaware of any benchmark setup in Indian languages. Hindi is one of the most widely spoken languages in the world. In this paper, we present an Event Extraction framework for Hindi language by creating an annotated resource for benchmarking, and then developing deep learning based models to set as the baselines. We crawl more than seventeen hundred disaster related Hindi news articles from the various news sources. We also develop deep learning based models for Event Trigger Detection and Classification, Argument Detection and Classification and Event-Argument Linking.
This paper proposes a representation framework for encoding spatial language in radiology based on frame semantics. The framework is adopted from the existing SpatialNet representation in the general domain with the aim to generate more accurate representations of spatial language used by radiologists. We describe Rad-SpatialNet in detail along with illustrating the importance of incorporating domain knowledge in understanding the varied linguistic expressions involved in different radiological spatial relations. This work also constructs a corpus of 400 radiology reports of three examination types (chest X-rays, brain MRIs, and babygrams) annotated with fine-grained contextual information according to this schema. Spatial trigger expressions and elements corresponding to a spatial frame are annotated. We apply BERT-based models (BERT-Base and BERT- Large) to first extract the trigger terms (lexical units for a spatial frame) and then to identify the related frame elements. The results of BERT- Large are decent, with F1 of 77.89 for spatial trigger extraction and an overall F1 of 81.61 and 66.25 across all frame elements using gold and predicted spatial triggers respectively. This frame-based resource can be used to develop and evaluate more advanced natural language processing (NLP) methods for extracting fine-grained spatial information from radiology text in the future.
Recent advances in neural computing and word embeddings for semantic processing open many new applications areas which had been left unaddressed so far because of inadequate language understanding capacity. But this new kind of approaches rely even more on training data to be operational. Corpora for financial applications exists, but most of them concern stock market prediction and are in English. To address this need for the French language and regulation oriented applications which require a deeper understanding of the text content, we hereby present “DoRe”, a French and dialectal French Corpus for NLP analytics in Finance, Regulation and Investment. This corpus is composed of: (a) 1769 Annual Reports from 336 companies among the most capitalized companies in: France (Euronext Paris) & Belgium (Euronext Brussels), covering a time frame from 2009 to 2019, and (b) related MetaData containing information for each company about its ISIN code, capitalization and sector. This corpus is designed to be as modular as possible in order to allow for maximum reuse in different tasks pertaining to Economics, Finance and Regulation. After presenting existing resources, we relate the construction of the DoRe corpus and the rationale behind our choices, concluding on the spectrum of possible uses of this new resource for NLP applications.
Brain signals are captured by clinical electroencephalography (EEG) which is an excellent tool for probing neural function. When EEG tests are performed, a textual EEG report is generated by the neurologist to document the findings, thus using language that describes the brain signals and its clinical correlations. Even with the impetus provided by the BRAIN initiative (brainitititive.nih.gov), there are no annotations available in texts that capture language describing the brain activities and their correlations with various pathologies. In this paper we describe an annotation effort carried out on a large corpus of EEG reports, providing examples of EEG-specific and clinically relevant concepts. In addition, we detail our annotation schema for brain signal attributes. We also discuss the resulting annotation of long-distance relations between concepts in EEG reports. By exemplifying a self-attention joint-learning to predict similar annotations in the EEG report corpus, we discuss the promising results, hoping that our effort will inform the design of novel knowledge capture techniques that will include the language of brain signals.
Artificial General Intelligence (AGI) is showing growing performance in numerous applications - beating human performance in Chess and Go, using knowledge bases and text sources to answer questions (SQuAD) and even pass human examination (Aristo project). In this paper, we describe the results of AI Journey, a competition of AI-systems aimed to improve AI performance on knowledge bases, reasoning and text generation. Competing systems pass the final native language exam (in Russian), including versatile grammar tasks (test and open questions) and an essay, achieving a high score of 69%, with 68% being an average human result. During the competition, a baseline for the task and essay parts was proposed, and 80+ systems were submitted, showing different approaches to task understanding and reasoning. All the data and solutions can be found on github https://github.com/sberbank-ai/combined_solution_aij2019
Creating ontologies is an expensive task. Our vision is that we can automatically generate ontologies based on a set of relevant documents to create a kick-start in ontology creating sessions. In this paper, we focus on enhancing two often used methods, OpenIE and co-occurrences. We evaluate the methods on two document sets, one about pizza and one about the agriculture domain. The methods are evaluated using two types of F1-score (objective, quantitative) and through a human assessment (subjective, qualitative). The results show that 1) Cooc performs both objectively and subjectively better than OpenIE; 2) the filtering methods based on keywords and on Word2vec perform similarly; 3) the filtering methods both perform better compared to OpenIE and similar to Cooc; 4) Cooc-NVP performs best, especially considering the subjective evaluation. Although, the investigated methods provide a good start for extracting an ontology out of a set of domain documents, various improvements are still possible, especially in the natural language based methods.
In financial services industry, compliance involves a series of practices and controls in order to meet key regulatory standards which aim to reduce financial risk and crime, e.g. money laundering and financing of terrorism. Faced with the growing risks, it is imperative for financial institutions to seek automated information extraction techniques for monitoring financial activities of their customers. This work describes an ontology of compliance-related concepts and relationships along with a corpus annotated according to it. The presented corpus consists of financial news articles in French and allows for training and evaluating domain-specific named entity recognition and relation extraction algorithms. We present some of our experimental results on named entity recognition and relation extraction using our annotated corpus. We aim to furthermore use the the proposed ontology towards construction of a knowledge base of financial relations.
Lexical semantic resources may be built using various approaches such as extraction from corpora, integration of the relevant pieces of knowledge from the pre-existing knowledge resources, and endogenous inference. Each of these techniques needs human supervision in order to deal with the potential errors, mapping difficulties or inferred candidate validation. We detail how various inference processes can be employed for the less supervised lexical semantic resource building. Our experience is based on the combination of different inference techniques for multilingual resource building and evaluation.
In this paper, we introduce our psychological approach to collect human-specific social knowledge from a text corpus, using NLP techniques. It is often not explicitly described but shared among people, which we call social knowledge. We focus on the social knowledge, especially personality and driving. We used the language resources that were developed based on psychological research methods; a Japanese personality dictionary (317 words) and a driving experience corpus (8,080 sentences) annotated with behavior and subjectivity. Using them, we automatically extracted collocations between personality descriptors and driving-related behavior from a driving behavior and subjectivity corpus (1,803,328 sentences after filtering) and obtained unique 5,334 collocations. To evaluate the collocations as social knowledge, we designed four step-by-step crowdsourcing tasks. They resulted in 266 pieces of social knowledge. They include the knowledge that might be difficult to recall by themselves but easy to agree with. We discuss the acquired social knowledge and the contribution to implementations into systems.
When speaking or writing, people omit information that seems clear and evident, such that only part of the message is expressed in words. Especially in argumentative texts it is very common that (important) parts of the argument are implied and omitted. We hypothesize that for argument analysis it will be beneficial to reconstruct this implied information. As a starting point for filling knowledge gaps, we build a corpus consisting of high-quality human annotations of missing and implied information in argumentative texts. To learn more about the characteristics of both the argumentative texts and the added information, we further annotate the data with semantic clause types and commonsense knowledge relations. The outcome of our work is a carefully designed and richly annotated dataset, for which we then provide an in-depth analysis by investigating characteristic distributions and correlations of the assigned labels. We reveal interesting patterns and intersections between the annotation categories and properties of our dataset, which enable insights into the characteristics of both argumentative texts and implicit knowledge in terms of structural features and semantic information. The results of our analysis can help to assist automated argument analysis and can guide the process of revealing implicit information in argumentative texts automatically.
We present MKGDB, a large-scale graph database created as a combination of multiple taxonomy backbones extracted from 5 existing knowledge graphs, namely: ConceptNet, DBpedia, WebIsAGraph, WordNet and the Wikipedia category hierarchy. MKGDB, thanks the versatility of the Neo4j graph database manager technology, is intended to favour and help the development of open-domain natural language processing applications relying on knowledge bases, such as information extraction, hypernymy discovery, topic clustering, and others. Our resource consists of a large hypernymy graph which counts more than 37 million nodes and more than 81 million hypernymy relations.
Legal technology is currently receiving a lot of attention from various angles. In this contribution we describe the main technical components of a system that is currently under development in the European innovation project Lynx, which includes partners from industry and research. The key contribution of this paper is a workflow manager that enables the flexible orchestration of workflows based on a portfolio of Natural Language Processing and Content Curation services as well as a Multilingual Legal Knowledge Graph that contains semantic information and meaningful references to legal documents. We also describe different use cases with which we experiment and develop prototypical solutions.
In this work, we address the task of extracting application-specific taxonomies from the category hierarchy of Wikipedia. Previous work on pruning the Wikipedia knowledge graph relied on silver standard taxonomies which can only be automatically extracted for a small subset of domains rooted in relatively focused nodes, placed at an intermediate level in the knowledge graphs. In this work, we propose an iterative methodology to extract an application-specific gold standard dataset from a knowledge graph and an evaluation framework to comparatively assess the quality of noisy automatically extracted taxonomies. We employ an existing state of the art algorithm in an iterative manner and we propose several sampling strategies to reduce the amount of manual work needed for evaluation. A first gold standard dataset is released to the research community for this task along with a companion evaluation framework. This dataset addresses a real-world application from the medical domain, namely the extraction of food-drug and herb-drug interactions.
Various research works have dealt with the comprehensibility of textual, audio, or audiovisual documents, and showed that factors related to text (e.g. linguistic complexity), sound (e.g. speech intelligibility), image (e.g. presence of visual context), or even to cognition and emotion can play a major role in the ability of humans to understand the semantic and pragmatic contents of a given document. However, to date, no reference human data is available that could help investigating the role of the linguistic and extralinguistic information present at these different levels (i.e., linguistic, audio/phonetic, and visual) in multimodal documents (e.g., movies). The present work aimed at building a corpus of human annotations that would help to study further how much and in which way the human perception of comprehensibility (i.e., of the difficulty of comprehension, referred in this paper as overall difficulty) of audiovisual documents is affected (1) by lexical complexity, grammatical complexity, and speech intelligibility, and (2) by the modality/ies (text, audio, video) available to the human recipient.
In scientific and technical communication, multiword terms are the most frequent type of lexical units. Rendering them in another language is not an easy task due to their cognitive complexity, the proliferation of different forms, and their unsystematic representation in terminographic resources. This often results in a broad spectrum of translations for multiword terms, which also foment term variation since they consist of two or more constituents. In this study we carried out a quantitative and qualitative analysis of Spanish translation variants of a set of environment-related concepts by evaluating equivalents in three parallel corpora, two comparable corpora and two terminological resources. Our results showed that MWTs exhibit a significant degree of term variation of different characteristics, which were used to establish a set of criteria according to which term variants should be selected, organized and described in terminological knowledge bases.
Recognizing spatial relations and reasoning about them is essential in multiple applications including navigation, direction giving and human-computer interaction in general. Spatial relations between objects can either be explicit – expressed as spatial prepositions, or implicit – expressed by spatial verbs such as moving, walking, shifting, etc. Both these, but implicit relations in particular, require significant common sense understanding. In this paper, we introduce the task of inferring implicit and explicit spatial relations between two entities in an image. We design a model that uses both textual and visual information to predict the spatial relations, making use of both positional and size information of objects and image embeddings. We contrast our spatial model with powerful language models and show how our modeling complements the power of these, improving prediction accuracy and coverage and facilitates dealing with unseen subjects, objects and relations.
Wikipedia is the largest web-based open encyclopedia covering more than three hundred languages. However, different language editions of Wikipedia differ significantly in terms of their information coverage. We present a systematic comparison of information coverage in English Wikipedia (most exhaustive) and Wikipedias in eight other widely spoken languages (Arabic, German, Hindi, Korean, Portuguese, Russian, Spanish and Turkish). We analyze the content present in the respective Wikipedias in terms of the coverage of topics as well as the depth of coverage of topics included in these Wikipedias. Our analysis quantifies and provides useful insights about the information gap that exists between different language editions of Wikipedia and offers a roadmap for the IR community to bridge this gap.
In this paper, we present Pártélet, a digitized Hungarian corpus of Communist propaganda texts. Pártélet was the official journal of the governing party during the Hungarian socialism from 1956 to 1989, hence it represents the direct political agitation and propaganda of the dictatorial system in question. The paper has a dual purpose: first, to present a general review of the corpus compilation process and the basic statistical data of the corpus, and second, to demonstrate through two case studies what the dataset can be used for. We show that our corpus provides a unique opportunity for conducting research on Hungarian propaganda discourse, as well as analyzing changes of this discourse over a 35-year period of time with computer-assisted methods.
A major domain of research in natural language processing is named entity recognition and disambiguation (NERD). One of the main ways of attempting to achieve this goal is through use of Semantic Web technologies and its structured data formats. Due to the nature of structured data, information can be extracted more easily, therewith allowing for the creation of knowledge graphs. In order to properly evaluate a NERD system, gold standard data sets are required. A plethora of different evaluation data sets exists, mostly relying on either Wikipedia or DBpedia. Therefore, we have extended a widely-used gold standard data set, KORE 50, to not only accommodate NERD tasks for DBpedia, but also for YAGO, Wikidata and Crunchbase. As such, our data set, KORE 50ˆDYWC, allows for a broader spectrum of evaluation. Among others, the knowledge graph agnosticity of NERD systems may be evaluated which, to the best of our knowledge, was not possible until now for this number of knowledge graphs.
Eye4Ref is a rich multimodal dataset of eye-movement recordings collected from referentially complex situated settings where the linguistic utterances and their visual referential world were available to the listener. It consists of not only fixation parameters but also saccadic movement parameters that are time-locked to accompanying German utterances (with English translations). Additionally, it also contains symbolic knowledge (contextual) representations of the images to map the referring expressions onto the objects in corresponding images. Overall, the data was collected from 62 participants in three different experimental setups (86 systematically controlled sentence–image pairs and 1844 eye-movement recordings). Referential complexity was controlled by visual manipulations (e.g. number of objects, visibility of the target items, etc.), and by linguistic manipulations (e.g., the position of the disambiguating word in a sentence). This multimodal dataset, in which the three different sources of information namely eye-tracking, language, and visual environment are aligned, offers a test of various research questions not from only language perspective but also computer vision.
Pre-trained models have achieved great success in learning unsupervised language representations by self-supervised tasks on large-scale corpora. Recent studies mainly focus on how to fine-tune different downstream tasks from a general pre-trained model. However, some studies show that customized self-supervised tasks for a particular type of downstream task can effectively help the pre-trained model to capture more corresponding knowledge and semantic information. Hence a new pre-training task called Sentence Insertion (SI) is proposed in this paper for Chinese query-passage pairs NLP tasks including answer span prediction, retrieval question answering and sentence level cloze test. The related experiment results indicate that the proposed SI can improve the performance of the Chinese Pre-trained models significantly. Moreover, a word segmentation method called SentencePiece is utilized to further enhance Chinese Bert performance for tasks with long texts. The complete source code is available at https://github.com/ewrfcas/SiBert_tensorflow.
This paper describes the Dakshina dataset, a new resource consisting of text in both the Latin and native scripts for 12 South Asian languages. The dataset includes, for each language: 1) native script Wikipedia text; 2) a romanization lexicon; and 3) full sentence parallel data in both a native script of the language and the basic Latin alphabet. We document the methods used for preparation and selection of the Wikipedia text in each language; collection of attested romanizations for sampled lexicons; and manual romanization of held-out sentences from the native script collections. We additionally provide baseline results on several tasks made possible by the dataset, including single word transliteration, full sentence transliteration, and language modeling of native script and romanized text.
We introduce the GM-RKB WikiText Error Correction Task for the automatic detection and correction of typographical errors in WikiText annotated pages. The included corpus is based on a snapshot of the GM-RKB domain-specific semantic wiki consisting of a large collection of concepts, personages, and publications primary centered on data mining and machine learning research topics. Numerous Wikipedia pages were also included as additional training data in the task’s evaluation process. The corpus was then automatically updated to synthetically include realistic errors to produce a training and evaluation ground truth comparison. We designed and evaluated two supervised baseline WikiFixer error correction methods: (1) a naive approach based on a maximum likelihood character-level language model; (2) and an advanced model based on a sequence-to-sequence (seq2seq) neural network architecture. Both error correction models operated at a character level. When compared against an off-the-shelf word-level spell checker these methods showed a significant improvement in the task’s performance – with the seq2seq-based model correcting a higher number of errors than it introduced. Finally, we published our data and code.
Prior work has determined domain similarity using text-based features of a corpus. However, when using pre-trained word embeddings, the underlying text corpus might not be accessible anymore. Therefore, we propose the CCA measure, a new measure of domain similarity based directly on the dimension-wise correlations between corresponding embedding spaces. Our results suggest that an inherent notion of domain can be captured this way, as we are able to reproduce our findings for different domain comparisons for English, German, Spanish and Czech as well as in cross-lingual comparisons. We further find a threshold at which the CCA measure indicates that two corpora come from the same domain in a monolingual setting by applying permutation tests. By evaluating the usability of the CCA measure in a domain adaptation application, we also show that it can be used to determine which corpora are more similar to each other in a cross-domain sentiment detection task.
We propose a new multilingual language model benchmark that is composed of 40+ languages spanning several scripts and linguistic families. With around 40 billion characters, we hope this new resource will accelerate the research of multilingual modeling. We train monolingual causal language models using a state-of-the-art model (Transformer-XL) establishing baselines for many languages. We also introduce the task of multilingual causal language modeling where we train our model on the combined text of 40+ languages from Wikipedia with different vocabulary sizes and evaluate on the languages individually. We released the cleaned-up text of 40+ Wikipedia language editions, the corresponding trained monolingual language models, and several multilingual language models with different fixed vocabulary sizes.
This paper proposes a novel framework for digital curation of Web corpora in order to provide robust estimation of their parameters, such as their composition and the lexicon. In recent years language models pre-trained on large corpora emerged as clear winners in numerous NLP tasks, but no proper analysis of the corpora which led to their success has been conducted. The paper presents a procedure for robust frequency estimation, which helps in establishing the core lexicon for a given corpus, as well as a procedure for estimating the corpus composition via unsupervised topic models and via supervised genre classification of Web pages. The results of the digital curation study applied to several Web-derived corpora demonstrate their considerable differences. First, this concerns different frequency bursts which impact the core lexicon obtained from each corpus. Second, this concerns the kinds of texts they contain. For example, OpenWebText contains considerably more topical news and political argumentation in comparison to ukWac or Wikipedia. The tools and the results of analysis have been released.
In this work, we consider the problem of personalizing language models, that is, building language models that are tailored to the writing style of an individual. Because training language models requires a large amount of text, and individuals do not necessarily possess a large corpus of their writing that could be used for training, approaches to personalizing language models must be able to rely on only a small amount of text from any one user. In this work, we compare three approaches to personalizing a language model that was trained on a large background corpus using a relatively small amount of text from an individual user. We evaluate these approaches using perplexity, as well as two measures based on next word prediction for smartphone soft keyboards. Our results show that when only a small amount of user-specific text is available, an approach based on priming gives the most improvement, while when larger amounts of user-specific text are available, an approach based on language model interpolation performs best. We carry out further experiments to show that these approaches to personalization outperform language model adaptation based on demographic factors.
In the paper, we present class-based LSTM Russian language models (LMs) with classes generated with the use of both word frequency and linguistic information data, obtained with the help of the “VisualSynan” software from the AOT project. We have created LSTM LMs with various numbers of classes and compared them with word-based LM and class-based LM with word2vec class generation in terms of perplexity, training time, and WER. In addition, we performed a linear interpolation of LSTM language models with the baseline 3-gram language model. The LSTM language models were used for very large vocabulary continuous Russian speech recognition at an N-best list rescoring stage. We achieved significant progress in training time reduction with only slight degradation in recognition accuracy comparing to the word-based LM. In addition, our LM with classes generated using linguistic information outperformed LM with classes generated using word2vec. We achieved WER of 14.94 % at our own speech corpus of continuous Russian speech that is 15 % relative reduction with respect to the baseline 3-gram model.
The recent success of pretrained language models in Natural Language Processing has sparked interest in training such models for languages other than English. Currently, training of these models can either be monolingual or multilingual based. In the case of multilingual models, such models are trained on concatenated data of multiple languages. We introduce AfriBERT, a language model for the Afrikaans language based on Bidirectional Encoder Representation from Transformers (BERT). We compare the performance of AfriBERT against multilingual BERT in multiple downstream tasks, namely part-of-speech tagging, named-entity recognition, and dependency parsing. Our results show that AfriBERT improves the current state-of-the-art in most of the tasks we considered, and that transfer learning from multilingual to monolingual model can have a significant performance improvement on downstream tasks. We release the pretrained model for AfriBERT.
Language models have become a key step to achieve state-of-the art results in many different Natural Language Processing (NLP) tasks. Leveraging the huge amount of unlabeled texts nowadays available, they provide an efficient way to pre-train continuous word representations that can be fine-tuned for a downstream task, along with their contextualization at the sentence level. This has been widely demonstrated for English using contextualized representations (Dai and Le, 2015; Peters et al., 2018; Howard and Ruder, 2018; Radford et al., 2018; Devlin et al., 2019; Yang et al., 2019b). In this paper, we introduce and share FlauBERT, a model learned on a very large and heterogeneous French corpus. Models of different sizes are trained using the new CNRS (French National Centre for Scientific Research) Jean Zay supercomputer. We apply our French language models to diverse NLP tasks (text classification, paraphrasing, natural language inference, parsing, word sense disambiguation) and show that most of the time they outperform other pre-training approaches. Different versions of FlauBERT as well as a unified evaluation protocol for the downstream tasks, called FLUE (French Language Understanding Evaluation), are shared to the research community for further reproducible experiments in French NLP.
Word clustering groups words that exhibit similar properties. One popular method for this is Brown clustering, which uses short-range distributional information to construct clusters. Specifically, this is a hard hierarchical clustering with a fixed-width beam that employs bi-grams and greedily minimizes global mutual information loss. The result is word clusters that tend to outperform or complement other word representations, especially when constrained by small datasets. However, Brown clustering has high computational complexity and does not lend itself to parallel computation. This, together with the lack of efficient implementations, limits their applicability in NLP. We present efficient implementations of Brown clustering and the alternative Exchange clustering as well as a number of methods to accelerate the computation of both hierarchical and flat clusters. We show empirically that clusters obtained with the accelerated method match the performance of clusters computed using the original methods.
This work aims to better understand the role of rhythm in foreign accent, and its modelling. We made a model of rhythm in French taking into account its variability, thanks to the Corpus pour l’Étude du Français Contemporain (CEFC), which contains up to 300 hours of speech of a wide variety of speaker profiles and situations. 16 parameters were computed, each of them being based on segment duration, such as voicing and intersyllabic timing. All the parameters are fully automatically detected from signal, without ASR or transcription. A gaussian mixture model was trained on 1,340 native speakers of French; any 30-second minimum speech may be computed to get the probability of its belonging to this model. We tested it with 146 test native speakers (NS), 37 non-native speakers (NNS) from the same corpus, and 29 non-native Japanese learners of French (JpNNS) from an independent corpus. The probability of NNS having inferior log-likelihood to NS was only a tendency (p=.067), maybe due to the heterogeneity of French proficiency of the speakers; but a much bigger probability was obtained for JpNNS (p<.0001), where all speakers were A2 level. Eta-squared test showed that most efficient parameters were intersyllabic mean duration and variation coefficient, along with speech rate for NNS; and speech rate and phonation ratio for JpNNS.
Terminological resources have proven crucial in many applications ranging from Computer-Aided Translation tools to authoring softwares and multilingual and cross-lingual information retrieval systems. Nonetheless, with the exception of a few felicitous examples, such as the IATE (Interactive Terminology for Europe) Termbank, many terminological resources are not available in standard formats, such as Term Base eXchange (TBX), thus preventing their sharing and reuse. Yet, these terminologies could be improved associating the correspondent ontology-based information. The research described in the present contribution demonstrates the process and the methodologies adopted in the automatic conversion into TBX of such type of resources, together with their semantic enrichment based on the formalization of ontological information into terminologies. We present a proof-of-concept using the Italian Linguistic Resource for the Archaeological domain (developed according to Thesauri and Guidelines of the Italian Central Institute for the Catalogue and Documentation). Further, we introduce the conversion tool developed to support the process of creating ontology-aware terminologies for improving interoperability and sharing of existing language technologies and data sets.
In this paper, we introduce an extension of the Berkeley FrameNet for the structured and semantic modeling of factual claims. Modeling is a robust tool that can be leveraged in many different tasks such as matching claims to existing fact-checks and translating claims to structured queries. Our work introduces 11 new manually crafted frames along with 9 existing FrameNet frames, all of which have been selected with fact-checking in mind. Along with these frames, we are also providing 2,540 fully annotated sentences, which can be used to understand how these frames are intended to work and to train machine learning models. Finally, we are also releasing our annotation tool to facilitate other researchers to make their own local extensions to FrameNet.
We introduce the first attempt at automatic speech recognition (ASR) in Inuktitut, as a representative for polysynthetic, low-resource languages, like many of the 900 Indigenous languages spoken in the Americas. As most previous work on Inuktitut, we use texts from parliament proceedings, but in addition we have access to 23 hours of transcribed oral stories. With this corpus, we show that Inuktitut displays a much higher degree of polysynthesis than other agglutinative languages usually considered in ASR, such as Finnish or Turkish. Even with a vocabulary of 1.3 million words derived from proceedings and stories, held-out stories have more than 60% of words out-of-vocabulary. We train bi-directional LSTM acoustic models, then investigate word and subword units, morphemes and syllables, and a deep neural network that finds word boundaries in subword sequences. We show that acoustic decoding using syllables decorated with word boundary markers results in the lowest word error rate.
While text corpora have been steadily increasing in overall size, even very large corpora are not designed to represent global population demographics. For example, recent work has shown that existing English gigaword corpora over-represent inner-circle varieties from the US and the UK. To correct implicit geographic and demographic biases, this paper uses country-level population demographics to guide the construction of gigaword web corpora. The resulting corpora explicitly match the ground-truth geographic distribution of each language, thus equally representing language users from around the world. This is important because it ensures that speakers of under-resourced language varieties (i.e., Indian English or Algerian French) are represented, both in the corpora themselves but also in derivative resources like word embeddings.
The task of fake news detection is to distinguish legitimate news articles that describe real facts from those which convey deceiving and fictitious information. As the fake news phenomenon is omnipresent across all languages, it is crucial to be able to efficiently solve this problem for languages other than English. A common approach to this task is supervised classification using features of various complexity. Yet supervised machine learning requires substantial amount of annotated data. For English and a small number of other languages, annotated data availability is much higher, whereas for the vast majority of languages, it is almost scarce. We investigate whether machine translation at its present state could be successfully used as an automated technique for annotated corpora creation and augmentation for fake news detection focusing on the English-Urdu language pair. We train a fake news classifier for Urdu on (1) the manually annotated dataset originally in Urdu and (2) the machine-translated version of an existing annotated fake news dataset originally in English. We show that at the present state of machine translation quality for the English-Urdu language pair, the fully automated data augmentation through machine translation did not provide improvement for fake news detection in Urdu.
Since word embeddings have been the most popular input for many NLP tasks, evaluating their quality is critical. Most research efforts are focusing on English word embeddings. This paper addresses the problem of training and evaluating such models for the Greek language. We present a new word analogy test set considering the original English Word2vec analogy test set and some specific linguistic aspects of the Greek language as well. Moreover, we create a Greek version of WordSim353 test collection for a basic evaluation of word similarities. Produced resources are available for download. We test seven word vector models and our evaluation shows that we are able to create meaningful representations. Last, we discover that the morphological complexity of the Greek language and polysemy can influence the quality of the resulting word embeddings.
We present our work towards a dataset of Mycenaean Linear B sequences gathered from the Mycenaean inscriptions written in the 13th and 14th century B.C. (c. 1400-1200 B.C.). The dataset contains sequences of Mycenaean words and ideograms according to the rules of the Mycenaean Greek language in the Late Bronze Age. Our ultimate goal is to contribute to the study, reading and understanding of ancient scripts and languages. Focusing on sequences, we seek to exploit the structure of the entire language, not just the Mycenaean vocabulary, to analyse sequential patterns. We use the dataset to experiment on estimating the missing symbols in damaged inscriptions.
The Inuktitut language, a member of the Inuit-Yupik-Unangan language family, is spoken across Arctic Canada and noted for its morphological complexity. It is an official language of two territories, Nunavut and the Northwest Territories, and has recognition in additional regions. This paper describes a newly released sentence-aligned Inuktitut–English corpus based on the proceedings of the Legislative Assembly of Nunavut, covering sessions from April 1999 to June 2017. With approximately 1.3 million aligned sentence pairs, this is, to our knowledge, the largest parallel corpus of a polysynthetic language or an Indigenous language of the Americas released to date. The paper describes the alignment methodology used, the evaluation of the alignments, and preliminary experiments on statistical and neural machine translation (SMT and NMT) between Inuktitut and English, in both directions.
This paper investigates the use of bilingual word embeddings for mining Hiligaynon translations of English words. There is very little research on Hiligaynon, an extremely low-resource language of Malayo-Polynesian origin with over 9 million speakers in the Philippines (we found just one paper). We use a publicly available Hiligaynon corpus with only 300K words, and match it with a comparable corpus in English. As there are no bilingual resources available, we manually develop a English-Hiligaynon lexicon and use this to train bilingual word embeddings. But we fail to mine accurate translations due to the small amount of data. To find out if the same holds true for a related language pair, we simulate the same low-resource setup on English to German and arrive at similar results. We then vary the size of the comparable English and German corpora to determine the minimum corpus size necessary to achieve competitive results. Further, we investigate the role of the seed lexicon. We show that with the same corpus size but with a smaller seed lexicon, performance can surpass results of previous studies. We release the lexicon of 1,200 English-Hiligaynon word pairs we created to encourage further investigation.
It has been widely admitted that morphological analysis is an important step in automated text processing for morphologically rich languages. Evenki is a language with rich morphology, therefore a morphological analyser is highly desirable for processing Evenki texts and developing applications for Evenki. Although two morphological analysers for Evenki have already been developed, they are able to analyse less than a half of the available Evenki corpora. The aim of this paper is to create a new morphological analyser for Evenki. It is implemented using the Helsinki Finite-State Transducer toolkit (HFST). The lexc formalism is used to specify the morphotactic rules, which define the valid orderings of morphemes in a word. Morphophonological alternations and orthographic rules are described using the twol formalism. The lexicon is extracted from available machine-readable dictionaries. Since a part of the corpora belongs to texts in Evenki dialects, a version of the analyser with relaxed rules is developed for processing dialectal features. We evaluate the analyser on available Evenki corpora and estimate precision, recall and F-score. We obtain coverage scores of between 61% and 87% on the available Evenki corpora.
Word embeddings have been successfully trained in many languages. However, both intrinsic and extrinsic metrics are variable across languages, especially for languages that depart significantly from English in morphology and orthography. This study focuses on building a word embedding model suitable for the Semitic language of Amharic (Ethiopia), which is both morphologically rich and written as an alphasyllabary (abugida) rather than an alphabet. We compare embeddings from tailored neural models, simple pre-processing steps, off-the-shelf baselines, and parallel tasks on a better-resourced Semitic language – Arabic. Experiments show our model’s performance on word analogy tasks, illustrating the divergent objectives of morphological vs. semantic analogies.
The use of the internet as a fast medium of spreading fake news reinforces the need for computational tools that combat it. Techniques that train fake news classifiers exist, but they all assume an abundance of resources including large labeled datasets and expert-curated corpora, which low-resource languages may not have. In this work, we make two main contributions: First, we alleviate resource scarcity by constructing the first expertly-curated benchmark dataset for fake news detection in Filipino, which we call “Fake News Filipino.” Second, we benchmark Transfer Learning (TL) techniques and show that they can be used to train robust fake news classifiers from little data, achieving 91% accuracy on our fake news dataset, reducing the error by 14% compared to established few-shot baselines. Furthermore, lifting ideas from multitask learning, we show that augmenting transformer-based transfer techniques with auxiliary language modeling losses improves their performance by adapting to writing style. Using this, we improve TL performance by 4-6%, achieving an accuracy of 96% on our best model. Lastly, we show that our method generalizes well to different types of news articles, including political news, entertainment news, and opinion articles.
Automatic analysis of connected speech by natural language processing techniques is a promising direction for diagnosing cognitive impairments. However, some difficulties still remain: the time required for manual narrative transcription and the decision on how transcripts should be divided into sentences for successful application of parsers used in metrics, such as Idea Density, to analyze the transcripts. The main goal of this paper was to develop a generic segmentation system for narratives of neuropsychological language tests. We explored the performance of our previous single-dataset-trained sentence segmentation architecture in a richer scenario involving three new datasets used to diagnose cognitive impairments, comprising different stories and two types of stimulus presentation for eliciting narratives — visual and oral — via illustrated story-book and sequence of scenes, and by retelling. Also, we proposed and evaluated three modifications to our previous RCNN architecture: (i) the inclusion of a Linear Chain CRF; (ii) the inclusion of a self-attention mechanism; and (iii) the replacement of the LSTM recurrent layer by a Quasi-Recurrent Neural Network layer. Our study allowed us to develop two new models for segmenting impaired speech transcriptions, along with an ideal combination of datasets and specific groups of narratives to be used as the training set.
Jejueo was classified as critically endangered by UNESCO in 2010. Although diverse efforts to revitalize it have been made, there have been few computational approaches. Motivated by this, we construct two new Jejueo datasets: Jejueo Interview Transcripts (JIT) and Jejueo Single Speaker Speech (JSS). The JIT dataset is a parallel corpus containing 170k+ Jejueo-Korean sentences, and the JSS dataset consists of 10k high-quality audio files recorded by a native Jejueo speaker and a transcript file. Subsequently, we build neural systems of machine translation and speech synthesis using them. All resources are publicly available via our GitHub repository. We hope that these datasets will attract interest of both language and machine learning communities.
Ainu is an unwritten language that has been spoken by Ainu people who are one of the ethnic groups in Japan. It is recognized as critically endangered by UNESCO and archiving and documentation of its language heritage is of paramount importance. Although a considerable amount of voice recordings of Ainu folklore has been produced and accumulated to save their culture, only a quite limited parts of them are transcribed so far. Thus, we started a project of automatic speech recognition (ASR) for the Ainu language in order to contribute to the development of annotated language archives. In this paper, we report speech corpus development and the structure and performance of end-to-end ASR for Ainu. We investigated four modeling units (phone, syllable, word piece, and word) and found that the syllable-based model performed best in terms of both word and phone recognition accuracy, which were about 60% and over 85% respectively in speaker-open condition. Furthermore, word and phone accuracy of 80% and 90% has been achieved in a speaker-closed setting. We also found out that a multilingual ASR training with additional speech corpora of English and Japanese further improves the speaker-open test accuracy.
This paper presents the development of a Guarani - Spanish parallel corpus with sentence-level alignment. The Guarani sentences of the corpus use the Jopara Guarani dialect, the dialect of Guarani spoken in Paraguay, which is based on Guarani grammar and may include several Spanish loanwords or neologisms. The corpus has around 14,500 sentence pairs aligned using a semi-automatic process, containing 228,000 Guarani tokens and 336,000 Spanish tokens extracted from web sources.
Automatic short answer grading is a significant problem in E-assessment. Several models have been proposed to deal with it. Evaluation and comparison of such solutions need the availability of Datasets with manual examples. In this paper, we introduce AR-ASAG, an Arabic Dataset for automatic short answer grading. The Dataset contains 2133 pairs of (Model Answer, Student Answer) in several versions (txt, xml, Moodle xml and .db). We explore then an unsupervised corpus based approach for automatic grading adapted to the Arabic Language. We use COALS (Correlated Occurrence Analogue to Lexical Semantic) algorithm to create semantic space for word distribution. The summation vector model is combined to term weighting and common words to achieve similarity between a teacher model answer and a student answer. The approach is particularly suitable for languages with scarce resources such as Arabic language where robust specific resources are not yet available. A set of experiments were conducted to analyze the effect of domain specificity, semantic space dimension and stemming techniques on the effectiveness of the grading model. The proposed approach gives promising results for Arabic language. The reported results may serve as baseline for future research work evaluation
Under-resourced and endangered or small languages yield problems for automatic processing and exploiting because of the small amount of available data. This paper shows an approach using different annotations of enriched linguistic research data to create communication boards commonly used in Alternative Augmentative Communication (AAC). Using manually created lexical analysis and rich annotation (instead of high data quantity) allows for an automated creation of AAC communication boards. The example presented in this paper uses data of the indigenous language Dolgan (an endangered Turkic language of Northern Siberia) created in the project INEL(Arkhipov and Däbritz, 2018) to generate a basic communication board with audio snippets to be used in e.g. hospital communication or for multilingual settings. The created boards can be importet into various AAC software. In addition, the usage of standard formats makes this approach applicable to various different use cases.
In this paper, we present a corpus of oral narratives from the Nisvai linguistic community and four associated language resources. Nisvai is an oral language spoken by 200 native speakers in the South-East of Malekula, an island of Vanuatu, Oceania. This language had never been the focus of a research before the one leading to this article. The corpus we present is made of 32 annotated narratives segmented into intonation units. The audio records were transcribed using the written conventions specifically developed for the language and translated into French. Four associated language resources have been generated by organizing the annotations into written documents: two of them are available online and two in paper format. The online resources allow the users to listen to the audio recordings whilereading the annotations. They were built to share the results of our fieldwork and to communicate on the Nisvai narrative practices with the researchers as well as with a more general audience. The bilingual paper resources, a booklet of narratives and a Nisvai-French French-Nisvai lexicon, were designed for the Nisvai community by taking into account their future uses (i.e. primary school).
Natural speech data on many languages have been collected by language documentation projects aiming to preserve lingustic and cultural traditions in audivisual records. These data hold great potential for large-scale cross-linguistic research into phonetics and language processing. Major obstacles to utilizing such data for typological studies include the non-homogenous nature of file formats and annotation conventions found both across and within archived collections. Moreover, time-aligned audio transcriptions are typically only available at the level of broad (multi-word) phrases but not at the word and segment levels. We report on solutions developed for these issues within the DoReCo (DOcumentation REference COrpus) project. DoReCo aims at providing time-aligned transcriptions for at least 50 collections of under-resourced languages. This paper gives a preliminary overview of the current state of the project and details our workflow, in particular standardization of formats and conventions, the addition of segmental alignments with WebMAUS, and DoReCo’s applicability for subsequent research programs. By making the data accessible to the scientific community, DoReCo is designed to bridge the gap between language documentation and linguistic inquiry.
Research in machine translation (MT) is developing at a rapid pace. However, most work in the community has focused on languages where large amounts of digital resources are available. In this study, we benchmark state of the art statistical and neural machine translation systems on two African languages which do not have large amounts of resources: Somali and Swahili. These languages are of social importance and serve as test-beds for developing technologies that perform reasonably well despite the low-resource constraint. Our findings suggest that statistical machine translation (SMT) and neural machine translation (NMT) can perform similarly in low-resource scenarios, but neural systems require more careful tuning to match performance. We also investigate how to exploit additional data, such as bilingual text harvested from the web, or user dictionaries; we find that NMT can significantly improve in performance with the use of these additional data. Finally, we survey the landscape of machine translation resources for the languages of Africa and provide some suggestions for promising future research directions.
St. Lawrence Island Yupik is an endangered polysynthetic language of the Bering Strait region. While conducting linguistic fieldwork between 2016 and 2019, we observed substantial support within the Yupik community for language revitalization and for resource development to support Yupik education. To that end, Chen & Schwartz (2018) implemented a finite-state morphological analyzer as a critical enabling technology for use in Yupik language education and technology. Chen & Schwartz (2018) reported a morphological analysis coverage rate of approximately 75% on a dataset of 60K Yupik tokens, leaving considerable room for improvement. In this work, we present a re-implementation of the Chen & Schwartz (2018) finite-state morphological analyzer for St. Lawrence Island Yupik that incorporates new linguistic insights; in particular, in this implementation we make use of the Paradigm Function Morphology (PFM) theory of morphology. We evaluate this new PFM-based morphological analyzer, and demonstrate that it consistently outperforms the existing analyzer of Chen & Schwartz (2018) with respect to accuracy and coverage rate across multiple datasets.
Zamboanga Chabacano (ZC) is the most vibrant variety of Philippine Creole Spanish, with over 400,000 native speakers in the Philippines (as of 2010). Following its introduction as a subject and a medium of instruction in the public schools of Zamboanga City from Grade 1 to 3 in 2012, an official orthography for this variety - the so-called “Zamboanga Chavacano Orthography” - has been approved in 2014. Its complexity, however, is a barrier to most speakers, since it does not necessarily reflect the particular phonetic evolution in ZC, but favours etymology instead. The distance between the correct spelling and the different spelling variations is often so great that delivering acceptable performance with the current de facto spell checking technologies may be challenging. The goals of this research have been to propose i) a spelling error taxonomy for ZC, formalised as an ontology and ii) an adaptive spell checking approach using Character-Based Statistical Machine Translation to correct spelling errors in ZC. Our results show that this approach is suitable for the goals mentioned and that it could be combined with other current spell checking technologies to achieve even higher performance.
We present in this paper our work on Algerian language, an under-resourced North African colloquial Arabic variety, for which we built a comparably large corpus of more than 36,000 code-switched user-generated comments annotated for sentiments. We opted for this data domain because Algerian is a colloquial language with no existing freely available corpora. Moreover, we compiled sentiment lexicons of positive and negative unigrams and bigrams reflecting the code-switches present in the language. We compare the performance of four models on the task of identifying sentiments, and the results indicate that a CNN model trained end-to-end fits better our unedited code-switched and unbalanced data across the predefined sentiment classes. Additionally, injecting the lexicons as background knowledge to the model boosts its performance on the minority class with a gain of 10.54 points on the F-score. The results of our experiments can be used as a baseline for future research for Algerian sentiment analysis.
This paper presents SwissCrawl, the largest Swiss German text corpus to date. Composed of more than half a million sentences, it was generated using a customized web scraping tool that could be applied to other low-resource languages as well. The approach demonstrates how freely available web pages can be used to construct comprehensive text corpora, which are of fundamental importance for natural language processing. In an experimental evaluation, we show that using the new corpus leads to significant improvements for the task of language modeling.
Cross-lingual word embeddings create a shared space for embeddings in two languages, and enable knowledge to be transferred between languages for tasks such as bilingual lexicon induction. One problem, however, is out-of-vocabulary (OOV) words, for which no embeddings are available. This is particularly problematic for low-resource and morphologically-rich languages, which often have relatively high OOV rates. Approaches to learning sub-word embeddings have been proposed to address the problem of OOV words, but most prior work has not considered sub-word embeddings in cross-lingual models. In this paper, we consider whether sub-word embeddings can be leveraged to form cross-lingual embeddings for OOV words. Specifically, we consider a novel bilingual lexicon induction task focused on OOV words, for language pairs covering several language families. Our results indicate that cross-lingual representations for OOV words can indeed be formed from sub-word embeddings, including in the case of a truly low-resource morphologically-rich language.
We introduce a dictionary containing normalized forms of common words in various Swiss German dialects into High German. As Swiss German is, for now, a predominantly spoken language, there is a significant variation in the written forms, even between speakers of the same dialect. To alleviate the uncertainty associated with this diversity, we complement the pairs of Swiss German - High German words with the Swiss German phonetic transcriptions (SAMPA). This dictionary becomes thus the first resource to combine large-scale spontaneous translation with phonetic transcriptions. Moreover, we control for the regional distribution and insure the equal representation of the major Swiss dialects. The coupling of the phonetic and written Swiss German forms is powerful. We show that they are sufficient to train a Transformer-based phoneme to grapheme model that generates credible novel Swiss German writings. In addition, we show that the inverse mapping - from graphemes to phonemes - can be modeled with a transformer trained with the novel dictionary. This generation of pronunciations for previously unknown words is key in training extensible automated speech recognition (ASR) systems, which are key beneficiaries of this dictionary.
The current situation regarding the existence of natural language processing (NLP) resources and tools for Corsican reveals their virtual non-existence. Our inventory contains only a few rare digital resources, lexical or corpus databases, requiring adaptation work. Our objective is to use the Banque de Données Langue Corse project (BDLC) to improve the availability of resources and tools for the Corsican language and, in the long term, provide a complete Basic Language Ressource Kit (BLARK). We have defined a roadmap setting out the actions to be undertaken: the collection of corpora and the setting up of a consultation interface (concordancer), and of a language detection tool, an electronic dictionary and a part-of-speech tagger. The first achievements regarding these topics have already been reached and are presented in this article. Some elements are also available on our project page (http://bdlc.univ-corse.fr/tal/).
Mi’kmaq is an Indigenous language spoken primarily in Eastern Canada. It is polysynthetic and low-resource. In this paper we consider a range of n-gram and RNN language models for Mi’kmaq. We find that an RNN language model, initialized with pre-trained fastText embeddings, performs best, highlighting the importance of sub-word information for Mi’kmaq language modelling. We further consider approaches to language modelling that incorporate cross-lingual word embeddings, but do not see improvements with these models. Finally we consider language models that operate over segmentations produced by SentencePiece — which include sub-word units as tokens — as opposed to word-level models. We see improvements for this approach over word-level language models, again indicating that sub-word modelling is important for Mi’kmaq language modelling.
This work introduces additions to the corpus ChoCo, a multimodal corpus for the American indigenous language Choctaw. Using texts from the corpus, we develop new computational resources by using two off-the-shelf tools: word2vec and Linguistica. Our work illustrates how these tools can be successfully implemented with a small corpus.
The success of several architectures to learn semantic representations from unannotated text and the availability of these kind of texts in online multilingual resources such as Wikipedia has facilitated the massive and automatic creation of resources for multiple languages. The evaluation of such resources is usually done for the high-resourced languages, where one has a smorgasbord of tasks and test sets to evaluate on. For low-resourced languages, the evaluation is more difficult and normally ignored, with the hope that the impressive capability of deep learning architectures to learn (multilingual) representations in the high-resourced setting holds in the low-resourced setting too. In this paper we focus on two African languages, Yorùbá and Twi, and compare the word embeddings obtained in this way, with word embeddings obtained from curated corpora and a language-dependent processing. We analyse the noise in the publicly available corpora, collect high quality and noisy data for the two languages and quantify the improvements that depend not only on the amount of data but on the quality too. We also use different architectures that learn word representations both from surface forms and characters to further exploit all the available information which showed to be important for these languages. For the evaluation, we manually translate the wordsim-353 word pairs dataset from English into Yorùbá and Twi. We extend the analysis to contextual word embeddings and evaluate multilingual BERT on a named entity recognition task. For this, we annotate with named entities the Global Voices corpus for Yorùbá. As output of the work, we provide corpora, embeddings and the test suits for both languages.
In this paper, we present and explain TRopBank “Turkish PropBank v2.0”. PropBank is a hand-annotated corpus of propositions which is used to obtain the predicate-argument information of a language. Predicate-argument information of a language can help understand semantic roles of arguments. “Turkish PropBank v2.0”, unlike PropBank v1.0, has a much more extensive list of Turkish verbs, with 17.673 verbs in total.
We present the Romanian legislative corpus which is a valuable linguistic asset for the development of machine translation systems, especially for under-resourced languages. The knowledge that can be extracted from this resource is necessary for a deeper understanding of how law terminology is used and how it can be made more consistent. At this moment the corpus contains more than 140k documents representing the legislative body of Romania. This corpus is processed and annotated at different levels: linguistically (tokenized, lemmatized and pos-tagged), dependency parsed, chunked, named entities identified and labeled with IATE terms and EUROVOC descriptors. Each annotated document has a CONLL-U Plus format consisting in 14 columns, in addition to the standard 10-column format, four other types of annotations were added. Moreover the repository will be periodically updated as new legislative texts are published. These will be automatically collected and transmitted to the processing and annotation pipeline. The access to the corpus will be done through ELRC infrastructure.
For most of the world’s languages, no primary data are available, even as many languages are disappearing. Throughout the last two decades, however, language documentation projects have produced substantial amounts of primary data from a wide variety of endangered languages. These resources are still in the early days of their exploration. One of the factors that makes them hard to use is a relative lack of standardized annotation conventions. In this paper, we will describe common practices in existing corpora in order to facilitate their future processing. After a brief introduction of the main formats used for annotation files, we will focus on commonly used tiers in the widespread ELAN and Toolbox formats. Minimally, corpora from language documentation contain a transcription tier and an aligned translation tier, which means they constitute parallel corpora. Additional common annotations include named references, morpheme separation, morpheme-by-morpheme glosses, part-of-speech tags and notes.
Given the lack of an annotated corpus of non-traditional Odia literature which serves as the standard when it comes sentiment analysis, we have created an annotated corpus of Odia sentences and made it publicly available to promote research in the field. Secondly, in order to test the usability of currently available Odia sentiment lexicon, we experimented with various classifiers by training and testing on the sentiment annotated corpus while using identified affective words from the same as features. Annotation and classification are done at sentence level as the usage of sentiment lexicon is best suited to sentiment analysis at this level. The created corpus contains 2045 Odia sentences from news domain annotated with sentiment labels using a well-defined annotation scheme. An inter-annotator agreement score of 0.79 is reported for the corpus.
This paper presents an approach for developing a task-oriented dialog system for less-resourced languages in scenarios where training data is not available. Both intent classification and slot filling are tackled. We project the existing annotations in rich-resource languages by means of Neural Machine Translation (NMT) and posterior word alignments. We then compare training on the projected monolingual data with direct model transfer alternatives. Intent Classifiers and slot filling sequence taggers are implemented using a BiLSTM architecture or by fine-tuning BERT transformer models. Models learnt exclusively from Basque projected data provide better accuracies for slot filling. Combining Basque projected train data with rich-resource languages data outperforms consistently models trained solely on projected data for intent classification. At any rate, we achieve competitive performance in both tasks, with accuracies of 81% for intent classification and 77% for slot filling.
In this paper, we report efforts towards the acquisition and construction of a bilingual parallel corpus between French and Wolof, a Niger-Congo language belonging to the Northern branch of the Atlantic group. The corpus is constructed as part of the SYSNET3LOc project. It currently contains about 70,000 French-Wolof parallel sentences drawn on various sources from different domains. The paper discusses the data collection procedure, conversion, and alignment of the corpus as well as it’s application as training data for neural machine translation. In fact, using this corpus, we were able to create word embedding models for Wolof with relatively good results. Currently, the corpus is being used to develop a neural machine translation model to translate French sentences into Wolof.
We present a new wordnet resource for Scottish Gaelic, a Celtic minority language spoken by about 60,000 speakers, most of whom live in Northwestern Scotland. The wordnet contains over 15 thousand word senses and was constructed by merging ten thousand new, high-quality translations, provided and validated by language experts, with an existing wordnet derived from Wiktionary. This new, considerably extended wordnet—currently among the 30 largest in the world—targets multiple communities: language speakers and learners; linguists; computer scientists solving problems related to natural language processing. By publishing it as a freely downloadable resource, we hope to contribute to the long-term preservation of Scottish Gaelic as a living language, both offline and on the Web.
Voice-based technologies are essential to cater to the hundreds of millions of new smartphone users. However, most of the languages spoken by these new users have little to no labelled speech data. Unfortunately, collecting labelled speech data in any language is an expensive and resource-intensive task. Moreover, existing platforms typically collect speech data only from urban speakers familiar with digital technology whose dialects are often very different from low-income users. In this paper, we explore the possibility of collecting labelled speech data directly from low-income workers. In addition to providing diversity to the speech dataset, we believe this approach can also provide valuable supplemental earning opportunities to these communities. To this end, we conducted a study where we collected labelled speech data in the Marathi language from three different user groups: low-income rural users, low-income urban users, and university students. Overall, we collected 109 hours of data from 36 participants. Our results show that the data collected from low-income participants is of comparable quality to the data collected from university students (who are typically employed to do this work) and that crowdsourcing speech data from low-income rural and urban workers is a viable method of gathering speech data.
We present the first resource focusing on the verbal inflectional morphology of San Juan Quiahije Chatino, a tonal mesoamerican language spoken in Mexico. We provide a collection of complete inflection tables of 198 lemmata, with morphological tags based on the UniMorph schema. We also provide baseline results on three core NLP tasks: morphological analysis, lemmatization, and morphological inflection.
The primary obstacle to developing technologies for low-resource languages is the lack of usable data. In this paper, we report the adaption and deployment of 4 technology-driven methods of data collection for Gondi, a low-resource vulnerable language spoken by around 2.3 million tribal people in south and central India. In the process of data collection, we also help in its revival by expanding access to information in Gondi through the creation of linguistic resources that can be used by the community, such as a dictionary, children’s stories, an app with Gondi content from multiple sources and an Interactive Voice Response (IVR) based mass awareness platform. At the end of these interventions, we collected a little less than 12,000 translated words and/or sentences and identified more than 650 community members whose help can be solicited for future translation efforts. The larger goal of the project is collecting enough data in Gondi to build and deploy viable language technologies like machine translation and speech to text systems that can help take the language onto the internet.
Irony is a linguistic device used to intend an idea while articulating an opposing expression. Many text analytic algorithms used for emotion extraction or sentiment analysis, produce invalid results due to the use of irony. Persian speakers use this device more often due to the language’s nature and some cultural reasons. This phenomenon also appears in social media platforms such as Twitter where users express their opinions using ironic or sarcastic posts. In the current research, which is the first attempt at irony detection in Persian language, emoji prediction is used to build a pretrained model. The model is finetuned utilizing a set of hand labeled tweets with irony tags. A bidirectional LSTM (BiLSTM) network is employed as the basis of our model which is improved by attention mechanism. Additionally, a Persian corpus for irony detection containing 4339 manually-labeled tweets is introduced. Experiments show the proposed approach outperforms the adapted state-of-the-art method tested on Persian dataset with an accuracy of 83.1%, and offers a strong baseline for further research in Persian language.
In this paper, we present computational resource grammars of Runyankore and Rukiga (R&R) languages. Runyankore and Rukiga are two under-resourced Bantu Languages spoken by about 6 million people indigenous to South- Western Uganda, East Africa. We used Grammatical Framework (GF), a multilingual grammar formalism and a special- purpose functional programming language to formalise the descriptive grammar of these languages. To the best of our knowledge, these computational resource grammars are the first attempt to the creation of language resources for R&R. In Future Work, we plan to use these grammars to bootstrap the generation of other linguistic resources such as multilingual corpora that make use of data-driven approaches to natural language processing feasible. In the meantime, they can be used to build Computer-Assisted Language Learning (CALL) applications for these languages among others.
Deep learning models are the current State-of-the-art methodologies towards many real-world problems. However, they need a substantial amount of labeled data to be trained appropriately. Acquiring labeled data can be challenging in some particular domains or less-resourced languages. There are some practical solutions regarding these issues, such as Active Learning and Transfer Learning. Active learning’s idea is simple: let the model choose the samples for annotation instead of labeling the whole dataset. This method leads to a more efficient annotation process. Active Learning models can achieve the baseline performance (the accuracy of the model trained on the whole dataset), with a considerably lower amount of labeled data. Several active learning approaches are tested in this work, and their compatibility with Persian is examined using a brand-new sentiment analysis dataset that is also introduced in this work. MirasOpinion, which to our knowledge is the largest Persian sentiment analysis dataset, is crawled from a Persian e-commerce website and annotated using a crowd-sourcing policy. LDA sampling, which is an efficient Active Learning strategy using Topic Modeling, is proposed in this research. Active Learning Strategies have shown promising results in the Persian language, and LDA sampling showed a competitive performance compared to other approaches.
Observing the damages that can be done by the rapid propagation of fake news in various sectors like politics and finance, automatic identification of fake news using linguistic analysis has drawn the attention of the research community. However, such methods are largely being developed for English where low resource languages remain out of the focus. But the risks spawned by fake and manipulative news are not confined by languages. In this work, we propose an annotated dataset of ≈ 50K news that can be used for building automated fake news detection systems for a low resource language like Bangla. Additionally, we provide an analysis of the dataset and develop a benchmark system with state of the art NLP techniques to identify Bangla fake news. To create this system, we explore traditional linguistic features and neural network based methods. We expect this dataset will be a valuable resource for building technologies to prevent the spreading of fake news and contribute in research with low resource languages. The dataset and source code are publicly available at https://github.com/Rowan1697/FakeNews.
We present a resource for computational experiments on Mapudungun, a polysynthetic indigenous language spoken in Chile with upwards of 200 thousand speakers. We provide 142 hours of culturally significant conversations in the domain of medical treatment. The conversations are fully transcribed and translated into Spanish. The transcriptions also include annotations for code-switching and non-standard pronunciations. We also provide baseline results on three core NLP tasks: speech recognition, speech synthesis, and machine translation between Spanish and Mapudungun. We further explore other applications for which the corpus will be suitable, including the study of code-switching, historical orthography change, linguistic structure, and sociological and anthropological studies.
Linguists seek insight from all human languages, however accessing information from most of the full store of extant global linguistic descriptions is not easy. One of the most common kinds of information that linguists have documented is vernacular sentences, as recorded in descriptive grammars. Typically these sentences are formatted as interlinear glossed text (IGT). Most descriptive grammars, however, exist only as hardcopy or scanned pdf documents. Consequently, parsing IGTs in scanned grammars is a priority, in order to significantly increase the volume of documented linguistic information that is readily accessible. Here we demonstrate fundamental viability for a technology that can assist in making a large number of linguistic data sources machine readable: the automated identification and parsing of interlinear glossed text from scanned page images. For example, we attain high median precision and recall (>0.95) in the identification of examples sentences in IGT format. Our results will be of interest to those who are keen to see more of the existing documentation of human language, especially for less-resourced and endangered languages, become more readily accessible.
We present findings from the creation of a massively parallel corpus in over 1600 languages, the Johns Hopkins University Bible Corpus (JHUBC). The corpus consists of over 4000 unique translations of the Christian Bible and counting. Our data is derived from scraping several online resources and merging them with existing corpora, combining them under a common scheme that is verse-parallel across all translations. We detail our effort to scrape, clean, align, and utilize this ripe multilingual dataset. The corpus captures the great typological variety of the world’s languages. We catalog this by showing highly similar proportions of representation of Ethnologue’s typological features in our corpus. We also give an example application: projecting pronoun features like clusivity across alignments to richly annotate languages which do not mark the distinction.
Since at least half of the world’s 6000 plus languages will vanish during the 21st century, language documentation has become a rapidly growing field in linguistics. A fundamental challenge for language documentation is the ”transcription bottleneck”. Speech technology may deliver the decisive breakthrough for overcoming the transcription bottleneck. This paper presents first experiments from the development of ASR4LD, a new automatic speech recognition (ASR) based tool for language documentation (LD). The experiments are based on recordings from an ongoing documentation project for the endangered Muyu language in New Guinea. We compare phoneme recognition experiments with American English, Austrian German and Slovenian as source language and Muyu as target language. The Slovenian acoustic models achieve the by far best performance (43.71% PER) in comparison to 57.14% PER with American English, and 89.49% PER with Austrian German. Whereas part of the errors can be explained by phonetic variation, the recording mismatch poses a major problem. On the long term, ASR4LD will not only be an integral part of the ongoing documentation project of Muyu, but will be further developed in order to facilitate also the language documentation process of other language groups.
This paper reports on challenges and solution approaches in the development of methods for language resource overarching data analysis in the field of language documentation. It is based on the successful outcomes of the initial phase of an 18 year long-term project on lesser resourced and mostly endangered indigenous languages of the Northern Eurasian area, which included the finalization and publication of multiple language corpora and additional language resources. While aiming at comprehensive cross-resource data analysis, the project at the same time is confronted with a dynamic and complex resource landscape, especially resulting from a vast amount of multi-layered information stored in the form of analogue primary data in different widespread archives on the territory of the Russian Federation. The methods described aim at solving the tension between unification of data sets and vocabularies on the one hand and maximum openness for the integration of future resources and adaption of external information on the other hand.
This work reports on the construction of a corpus of connected spoken Hong Kong Cantonese. The corpus aims at providing an additional resource for the study of modern (Hong Kong) Cantonese and also involves several controlled elicitation tasks which will serve different projects related to the phonology and semantics of Cantonese. The word-segmented corpus offers recordings, phonemic transcription, and Chinese characters transcription. The corpus contains a total of 768 minutes of recordings and transcripts of forty speakers. All the audio material has been aligned at utterance level with the transcriptions, using the ELAN transcription and annotation tool. The controlled elicitation task was based on the design of HCRC MapTask corpus (Anderson et al., 1991), in which participants had to communicate using solely verbal means as eye contact was restricted. In this paper, we outline the design of the maps and their landmarks and the basic segmentation principles of the data and various transcription conventions we adopted. We also compare the contents of Cantomap to those of comparable Cantonese corpora.
We introduce new monolingual corpora for four indigenous and endangered languages from Peru: Shipibo-konibo, Ashaninka, Yanesha and Yine. Given the total absence of these languages in the web, the extraction and processing of texts from PDF files is relevant in a truly low-resource language scenario. Our procedure for monolingual corpus creation considers language-specific and language-agnostic steps, and focuses on educational PDF files with multilingual sentences, noisy pages and low-structured content. Through an evaluation based on language modelling and character-level perplexity on a subset of manually extracted sentences, we determine that our method allows the creation of clean corpora for the four languages, a key resource for natural language processing tasks nowadays.
Making the low-resource language, Icelandic, accessible and usable in Language Technology is a work in progress and is supported by the Icelandic government. Creating resources and suitable training data (e.g., a dependency treebank) is a fundamental part of that work. We describe work on a parallel Icelandic dependency treebank based on Universal Dependencies (UD). This is important because it is the first parallel treebank resource for the language and since several other languages already have a resource based on the same text. Two Icelandic treebanks based on phrase-structure grammar have been built and ongoing work aims to convert them to UD. Previously, limited work has been done on dependency grammar for Icelandic. The current project aims to ameliorate this situation by creating a small dependency treebank from scratch. Creating a treebank is a laborious task so the process was implemented in an accessible manner using freely available tools and resources. The parallel data in the UD project was chosen as a source because this would furthermore give us the first parallel treebank for Icelandic. The Icelandic parallel UD corpus will be published as part of UD version 2.6.
This paper outlines the ongoing effort of creating the first treebank for Occitan, a low-ressourced regional language spoken mainly in the south of France. We briefly present the global context of the project and report on its current status. We adopt the Universal Dependencies framework for this project. Our methodology is based on two main principles. Firstly, in order to guarantee the annotation quality, we use the agile annotation approach. Secondly, we rely on pre-processing using existing tools (taggers and parsers) to facilitate the work of human annotators, mainly through a delexicalized cross-lingual parsing approach. We present the results available at this point (annotation guidelines and a sub-corpus annotated with PoS tags and lemmas) and give the timeline for the rest of the work.
This paper discusses the construction and the ongoing development of the Old Javanese Wordnet. The words were extracted from the digitized version of the Old Javanese–English Dictionary (Zoetmulder, 1982). The wordnet is built using the ‘expansion’ approach (Vossen, 1998), leveraging on the Princeton Wordnet’s core synsets and semantic hierarchy, as well as scientific names. The main goal of our project was to produce a high quality, human-curated resource. As of December 2019, the Old Javanese Wordnet contains 2,054 concepts or synsets and 5,911 senses. It is released under a Creative Commons Attribution 4.0 International License (CC BY 4.0). We are still developing it and adding more synsets and senses. We believe that the lexical data made available by this wordnet will be useful for a variety of future uses such as the development of Modern Javanese Wordnet and many language processing tasks and linguistic research on Javanese.
Mexico is a Spanish speaking country that has a great language diversity, with 68 linguistic groups and 364 varieties. As they face a lack of representation in education, government, public services and media, they present high levels of endangerment. Due to the lack of data available on social media and the internet, few technologies have been developed for these languages. To analyze different linguistic phenomena in the country, the Language Engineering Group developed the Corpus Paralelo de Lenguas Mexicanas (CPLM) [The Mexican Languages Parallel Corpus], a collaborative parallel corpus for the low-resourced languages of Mexico. The CPLM aligns Spanish with six indigenous languages: Maya, Ch’ol, Mazatec, Mixtec, Otomi, and Nahuatl. First, this paper describes the process of building the CPLM: text searching, digitalization and alignment process. Furthermore, we present some difficulties regarding dialectal and orthographic variations. Second, we present the interface and types of searching as well as the use of filters.
We introduce the SiNER: a named entity recognition (NER) dataset for low-resourced Sindhi language with quality baselines. It contains 1,338 news articles and more than 1.35 million tokens collected from Kawish and Awami Awaz Sindhi newspapers using the begin-inside-outside (BIO) tagging scheme. The proposed dataset is likely to be a significant resource for statistical Sindhi language processing. The ultimate goal of developing SiNER is to present a gold-standard dataset for Sindhi NER along with quality baselines. We implement several baseline approaches of conditional random field (CRF) and recent popular state-of-the-art bi-directional long-short term memory (Bi-LSTM) models. The promising F1-score of 89.16 outputted by the Bi-LSTM-CRF model with character-level representations demonstrates the quality of our proposed SiNER dataset.
The study of predicate frame is an important topic for semantic analysis. Abstract Meaning Representation (AMR) is an emerging graph based semantic representation of a sentence. Since core semantic roles defined in the predicate lexicon compose the backbone in an AMR graph, the construction of the lexicon becomes the key issue. The existing lexicons blur senses and frames of predicates, which needs to be refined to meet the tasks like word sense disambiguation and event extraction. This paper introduces the on-going project on constructing a novel predicate lexicon for Chinese AMR corpus. The new lexicon includes 14,389 senses and 10,800 frames of 8,470 words. As some senses can be aligned to more than one frame, and vice versa, we found the alignment between senses is not just one frame per sense. Explicit analysis is given for multiple aligned relations, which proves the necessity of the proposed lexicon for AMR corpus, and supplies real data for linguistic theoretical studies.
Multi-word expressions (MWEs) are a hot topic in research in natural language processing (NLP), including topics such as MWE detection, MWE decomposition, and research investigating the exploitation of MWEs in other NLP fields such as Machine Translation. However, the availability of bilingual or multi-lingual MWE corpora is very limited. The only bilingual MWE corpora that we are aware of is from the PARSEME (PARSing and Multi-word Expressions) EU Project. This is a small collection of only 871 pairs of English-German MWEs. In this paper, we present multi-lingual and bilingual MWE corpora that we have extracted from root parallel corpora. Our collections are 3,159,226 and 143,042 bilingual MWE pairs for German-English and Chinese-English respectively after filtering. We examine the quality of these extracted bilingual MWEs in MT experiments. Our initial experiments applying MWEs in MT show improved translation performances on MWE terms in qualitative analysis and better general evaluation scores in quantitative analysis, on both German-English and Chinese-English language pairs. We follow a standard experimental pipeline to create our MultiMWE corpora which are available online. Researchers can use this free corpus for their own models or use them in a knowledge base as model features.
Transliteration is generally a phonetically based transcription across different writing systems. It is a crucial task for various downstream natural language processing applications. For the Myanmar (Burmese) language, robust automatic transliteration for borrowed English words is a challenging task because of the complex Myanmar writing system and the lack of data. In this study, we constructed a Myanmar-English named entity dictionary containing more than eighty thousand transliteration instances. The data have been released under a CC BY-NC-SA license. We evaluated the automatic transliteration performance using statistical and neural network-based approaches based on the prepared data. The neural network model outperformed the statistical model significantly in terms of the BLEU score on the character level. Different units used in the Myanmar script for processing were also compared and discussed.
Embedding commonsense knowledge is crucial for end-to-end models to generalize inference beyond training corpora. However, existing word analogy datasets have tended to be handcrafted, involving permutations of hundreds of words with only dozens of pre-defined relations, mostly morphological relations and named entities. In this work, we model commonsense knowledge down to word-level analogical reasoning by leveraging E-HowNet, an ontology that annotates 88K Chinese words with their structured sense definitions and English translations. We present CA-EHN, the first commonsense word analogy dataset containing 90,505 analogies covering 5,656 words and 763 relations. Experiments show that CA-EHN stands out as a great indicator of how well word representations embed commonsense knowledge. The dataset is publicly available at https://github.com/ckiplab/CA-EHN.
Word senses are typically defined with textual definitions for human consumption and, in computational lexicons, put in context via lexical-semantic relations such as synonymy, antonymy, hypernymy, etc. In this paper we embrace a radically different paradigm that provides a slot-filler structure, called “semagram”, to define the meaning of words in terms of their prototypical semantic information. We propose a semagram-based knowledge model composed of 26 semantic relationships which integrates features from a range of different sources, such as computational lexicons and property norms. We describe an annotation exercise regarding 50 concepts over 10 different categories and put forward different automated approaches for extending the semagram base to thousands of concepts. We finally evaluated the impact of the proposed resource on a semantic similarity task, showing significant improvements over state-of-the-art word embeddings.
Cognate words, defined as words in different languages which derive from a common etymon, can be useful for language learners, who can leverage the orthographical similarity of cognates to more easily understand a text in a foreign language. Deceptive cognates, or false friends, do not share the same meaning anymore; these can be instead deceiving and detrimental for language acquisition or text understanding in a foreign language. We use an automatic method of detecting false friends from a set of cognates, in a fully unsupervised fashion, based on cross-lingual word embeddings. We implement our method for English and five Romance languages, including a low-resource language (Romanian), and evaluate it against two different gold standards. The method can be extended easily to any language pair, requiring only large monolingual corpora for the involved languages and a small bilingual dictionary for the pair. We additionally propose a measure of “falseness” of a false friends pair. We publish freely the database of false friends in the six languages, along with the falseness scores for each cognate pair. The resource is the largest of the kind that we are aware of, both in terms of languages covered and number of word pairs.
We present the parallel creation of a WordNet resource for Swedish and Bulgarian which is tightly aligned with the Princeton WordNet. The alignment is not only on the synset level, but also on word level, by matching words with their closest translations in each language. We argue that the tighter alignment is essential in machine translation and natural language generation. About one-fifth of the lexical entries are also linked to the corresponding Wikipedia articles. In addition to the traditional semantic relations in WordNet, we also integrate morphological and morpho-syntactic information. The resource comes with a corpus where examples from Princeton WordNet are translated to Swedish and Bulgarian. The examples are aligned on word and phrase level. The new resource is open-source and in its development we used only existing open-source resources.
This paper introduces ENGLAWI, a large, versatile, XML-encoded machine-readable dictionary extracted from Wiktionary. ENGLAWI contains 752,769 articles encoding the full body of information included in Wiktionary: simple words, compounds and multiword expressions, lemmas and inflectional paradigms, etymologies, phonemic transcriptions in IPA, definition glosses and usage examples, translations, semantic and morphological relations, spelling variants, etc. It is fully documented, released under a free license and supplied with G-PeTo, a series of scripts allowing easy information extraction from ENGLAWI. Additional resources extracted from ENGLAWI, such as an inflectional lexicon, a lexicon of diatopic variants and the inclusion dates of headwords in Wiktionary’s nomenclature are also provided. The paper describes the content of the resource and illustrates how it can be - and has been - used in previous studies. We finally introduce an ongoing work that computes lexicographic word embeddings from ENGLAWI’s definitions.
We introduce the Romance Verbal Inflection Dataset 2.0, a multilingual lexicon of Romance inflection covering 74 varieties. The lexicon provides verbal paradigm forms in broad IPA phonemic notation. Both lexemes and paradigm cells are organized to reflect cognacy. Such multi-lingual inflected lexicons annotated for two dimensions of cognacy are necessary to study the evolution of inflectional paradigms, and test linguistic hypotheses systematically. However, these resources seldom exist, and when they do, they are not usually encoded in computationally usable ways. The Oxford Online Database of Romance Verb Morphology provides this kind of information, however, it is not maintained anymore and is only available as a web service without interfaces for machine-readability. We collect its data and clean and correct it for consistency using both heuristics and expert annotator judgements. Most resources used to study language evolution computationally rely strictly on multilingual contemporary information, and lack information about prior stages of the languages. To provide such information, we augment the database with Latin paradigms from the LatInFlexi lexicon. Finally, to make it widely avalable, the resource is released under a GPLv3 license in CLDF format.
We present word2word, a publicly available dataset and an open-source Python package for cross-lingual word translations extracted from sentence-level parallel corpora. Our dataset provides top-k word translations in 3,564 (directed) language pairs across 62 languages in OpenSubtitles2018 (Lison et al., 2018). To obtain this dataset, we use a count-based bilingual lexicon extraction model based on the observation that not only source and target words but also source words themselves can be highly correlated. We illustrate that the resulting bilingual lexicons have high coverage and attain competitive translation quality for several language pairs. We wrap our dataset and model in an easy-to-use Python library, which supports downloading and retrieving top-k word translations in any of the supported language pairs as well as computing top-k word translations for custom parallel corpora.
The evaluation and exchange of large lexicon databases remains a challenge in many NLP applications. Despite the existence of commonly accepted standards for the format and the features used in a lexicon, there is still a lack of precise and interoperable specification requirements about how lexical entries of a particular language should look like, both in terms of the numbers of forms and in terms of features associated with these forms. This paper presents the notion of “lexical masks”, a powerful tool used to evaluate and exchange lexicon databases in many languages.
We present a large-scale 26,000-lemma leveled readability lexicon for Modern Standard Arabic. The lexicon was manually annotated in triplicate by language professionals from three regions in the Arab world. The annotations show a high degree of agreement; and major differences were limited to regional variations. Comparing lemma readability levels with their frequencies provided good insights in the benefits and pitfalls of frequency-based readability approaches. The lexicon will be publicly available.
Historical dictionaries of the pre-digital period are important resources for the study of older languages. Taking the example of the ‘Altfranzösisches Wörterbuch’, an Old French dictionary published from 1925 onwards, this contribution shows how the printed dictionaries can be turned into a more easily accessible and more sustainable lexical database, even though a full-text retro-conversion is too costly. Over 57,000 German sense definitions were identified in uncorrected OCR output. For verbs and nouns, 34,000 senses of more than 20,000 lemmas were matched with GermaNet, a semantic network for German, and, in a second step, linked to synsets of the English WordNet. These results are relevant for the automatic processing of Old French, for the annotation and exploitation of Old French text corpora, and for the philological study of Old French in general.
This paper introduces Cifu, a lexical database for Hong Kong Cantonese (HKC) that offers phonological and orthographic information, frequency measures, and lexical neighborhood information for lexical items in HKC. Cifu is of use for NLP applications and the design and analysis of psycholinguistics experiments on HKC. We elaborate on the characteristics and challenges specific to HKC that were relevant in the design of Cifu. This includes lexical, orthographic and phonological aspects of HKC, word segmentation issues, the place of HKC in written media, and the availability of data. We discuss the measure of Neighborhood Density (ND), highlighting how the analytic nature of Cantonese and its writing system affect that measure. We justify using six different variations of ND, based on the possibility of inserting or deleting phonemes when searching for neighbors and on the choice of data for retrieving frequencies. Statistics about the four genres (written, adult spoken, children spoken and child-directed) within the dataset are discussed. We find that the lexical diversity of the child-directed speech genre is particularly low, compared to a size-matched written corpus. The correlations of word frequencies of different genres are all high, but in generally decrease as word length increases.
Sentiment lexicons are essential for developing automatic sentiment analysis systems, but the resources currently available mostly cover modern languages. Lexicons for ancient languages are few and not evaluated with high-quality gold standards. However, the study of attitudes and emotions in ancient texts is a growing field of research which poses specific issues (e.g., lack of native speakers, limited amount of data, unusual textual genres for the sentiment analysis task, such as philosophical or documentary texts) and can have an impact on the work of scholars coming from several disciplines besides computational linguistics, e.g. historians and philologists. The work presented in this paper aims at providing the research community with a set of sentiment lexicons built by taking advantage of manually-curated resources belonging to the long tradition of Latin corpora and lexicons creation. Our interdisciplinary approach led us to release: i) two automatically generated sentiment lexicons; ii) a gold standard developed by two Latin language and culture experts; iii) a silver standard in which semantic and derivational relations are exploited so to extend the list of lexical items of the gold standard. In addition, the evaluation procedure is described together with a first application of the lexicons to a Latin tragedy.
There is a growing body of work on how word meaning changes over time: mutation. In contrast, there is very little work on how different words compete to represent the same meaning, and how the degree of success of words in that competition changes over time: natural selection. We present a new dataset, WordWars, with historical frequency data from the early 1800s to the early 2000s for monosemous English words in over 5000 synsets. We explore three broad questions with the dataset: (1) what is the degree to which predominant words in these synsets have changed, (2) how do prominent word features such as frequency, length, and concreteness impact natural selection, and (3) what are the differences between the predominant words of the 2000s and the predominant words of early 1800s. We show that close to one third of the synsets undergo a change in the predominant word in this time period. Manual annotation of these pairs shows that about 15% of these are orthographic variations, 25% involve affix changes, and 60% have completely different roots. We find that frequency, length, and concreteness all impact natural selection, albeit in different ways.
Cognates are present in multiple variants of the same text across different languages (e.g., “hund” in German and “hound” in the English language mean “dog”). They pose a challenge to various Natural Language Processing (NLP) applications such as Machine Translation, Cross-lingual Sense Disambiguation, Computational Phylogenetics, and Information Retrieval. A possible solution to address this challenge is to identify cognates across language pairs. In this paper, we describe the creation of two cognate datasets for twelve Indian languages namely Sanskrit, Hindi, Assamese, Oriya, Kannada, Gujarati, Tamil, Telugu, Punjabi, Bengali, Marathi, and Malayalam. We digitize the cognate data from an Indian language cognate dictionary and utilize linked Indian language Wordnets to generate cognate sets. Additionally, we use the Wordnet data to create a False Friends’ dataset for eleven language pairs. We also evaluate the efficacy of our dataset using previously available baseline cognate detection approaches. We also perform a manual evaluation with the help of lexicographers and release the curated gold-standard dataset with this paper.
We propose a new approach to constructing a personality dictionary with psychological evidence. In this study, we collect personality words, using word embeddings, and construct a personality dictionary with weights for Big Five traits. The weights are calculated based on the responses of the large sample (N=1,938, female = 1,004, M=49.8years old:20-78, SD=16.3). All the respondents answered a 20-item personality questionnaire and 537 personality items derived from word embeddings. We present the procedures to examine the qualities of responses with psychological methods and to calculate the weights. These result in a personality dictionary with two sub-dictionaries. We also discuss an application of the acquired resources.
Hedging is a commonly used strategy in conversational management to show the speaker’s lack of commitment to what they communicate, which may signal problems between the speakers. Our project is interested in examining the presence of hedging words and phrases in identifying the tension between an interviewer and interviewee during a survivor interview. While there have been studies on hedging detection in the natural language processing literature, all existing work has focused on structured texts and formal communications. Our project thus investigated a corpus of eight unstructured conversational interviews about the Rwanda Genocide and identified hedging patterns in the interviewees’ responses. Our work produced three manually constructed lists of hedge words, booster words, and hedging phrases. Leveraging these lexicons, we developed a rule-based algorithm that detects sentence-level hedges in informal conversations such as survivor interviews. Our work also produced a dataset of 3000 sentences having the categories Hedge and Non-hedge annotated by three researchers. With experiments on this annotated dataset, we verify the efficacy of our proposed algorithm. Our work contributes to the further development of tools that identify hedges from informal conversations and discussions.
We introduce three language resources for Japanese lexical simplification: 1) a large-scale word complexity lexicon, 2) the first synonym lexicon for converting complex words to simpler ones, and 3) the first toolkit for developing and benchmarking Japanese lexical simplification system. Our word complexity lexicon is expanded to a broader vocabulary using a classifier trained on a small, high-quality word complexity lexicon created by Japanese language teachers. Based on this word complexity estimator, we extracted simplified word pairs from a large-scale synonym lexicon and constructed a simplified synonym lexicon useful for lexical simplification. In addition, we developed a Python library that implements automatic evaluation and key methods in each subtask to ease the construction of a lexical simplification pipeline. Experimental results show that the proposed method based on our lexicon achieves the highest performance of Japanese lexical simplification. The current lexical simplification is mainly studied in English, which is rich in language resources such as lexicons and toolkits. The language resources constructed in this study will help advance the lexical simplification system in Japanese.
Given the well-established usefulness of part-of-speech tag annotations in many syntactically oriented downstream NLP tasks, the recently proposed notion of semantic tagging (Bjerva et al. 2016) aims at tagging words with tags informed by semantic distinctions, which are likely to be useful across a range of semantic tasks. To this end, their annotation scheme distinguishes, for instance, privative attributes from subsective ones. While annotated corpora exist, their size is limited and thus many words are out-of-vocabulary words. In this paper, we study to what extent we can automatically predict the tags associated with unseen words. We draw on large-scale word representation data to derive a large new Semantic Tag lexicon. Our experiments show that we can infer semantic tags for words with high accuracy both monolingually and cross-lingually.
LexiDB is a tool for storing, managing and querying corpus data. In contrast to other database management systems (DBMSs), it is designed specifically for text corpora. It improves on other corpus management systems (CMSs) because data can be added and deleted from corpora on the fly with the ability to add live data to existing corpora. LexiDB sits between these two categories of DBMSs and CMSs, more specialised to language data than a general purpose DBMS but more flexible than a traditional static corpus management system. Previous work has demonstrated the scalability of LexiDB in response to the growing need to be able to scale out for ever growing corpus datasets. Here, we present the patterns and methods developed in LexiDB for storage, retrieval and querying of multi-level annotated corpus data. These techniques are evaluated and compared to an existing CMS (Corpus Workbench CWB - CQP) and indexer (Lucene). We find that LexiDB consistently outperforms existing tools for corpus queries. This is particularly apparent with large corpora and when handling queries with large result sets
Valency lexicons usually describe valency behavior of verbs in non-reflexive and non-reciprocal constructions. However, reflexive and reciprocal constructions are common morphosyntactic forms of verbs. Both of these constructions are characterized by regular changes in morphosyntactic properties of verbs, thus they can be described by grammatical rules. On the other hand, the possibility to create reflexive and/or reciprocal constructions cannot be trivially derived from the morphosyntactic structure of verbs as it is conditioned by their semantic properties as well. A large-coverage valency lexicon allowing for rule based generation of all well formed verb constructions should thus integrate the information on reflexivity and reciprocity. In this paper, we propose a semi-automatic procedure, based on grammatical constraints on reflexivity and reciprocity, detecting those verbs that form reflexive and reciprocal constructions in corpus data. However, exploitation of corpus data for this purpose is complicated due to the diverse functions of reflexive markers crossing the domain of reflexivity and reciprocity. The list of verbs identified by the previous procedure is thus further used in an automatic experiment, applying word embeddings for detecting semantically similar verbs. These candidate verbs have been manually verified and annotation of their reflexive and reciprocal constructions has been integrated into the valency lexicon of Czech verbs VALLEX.
Classical Armenian is a poorly endowed language, that despite a great tradition of lexicographical erudition is coping with a lack of resources. Although numerous initiatives exist to preserve the Classical Armenian language, the lack of precise and complete grammatical and lexicographical resources remains. This article offers a situation analysis of the existing resources for Classical Armenian and presents the new digital resources provided on the Calfa platform. The Calfa project gathers existing resources and updates, enriches and enhances their content to offer the richest database for Classical Armenian today. Faced with the challenges specific to a poorly endowed language, the Calfa project is also developing new technologies and solutions to enable preservation, advanced research, and larger systems and developments for the Armenian language
As part of constructing the NINJAL Parsed Corpus of Modern Japanese (NPCMJ), a web-accessible language resource, we are adding frame information for predicates, together with two types of semantic role labels that mark the contributions of arguments. One role type consists of numbered semantic roles, like in PropBank, to capture relations between arguments in different syntactic patterns. The other role type consists of semantic roles with conventional names.Both role types are compatible with hierarchical frames that belong to related predicates. Adding semantic role and frame information to the NPCMJ will support a web environment where language learners and linguists can search examples of Japanese for syntactic and semantic features. The annotation will also provide a language resource for NLP researchers making semantic parsing models (e.g., for AMR parsing) following machine learning approaches. In this paper, we describe how the two types of semantic role labels are defined under the frame based approach, i.e., both types can be consistently applied when linked to corresponding frames. Then we show special cases of syntactic patterns and the current status of the annotation work.
In this article, we lay out the basic ideas and principles of the project Framing Situations in the Dutch Language. We provide our first results of data acquisition, together with the first data release. We introduce the notion of cross-lingual referential corpora. These corpora consist of texts that make reference to exactly the same incidents. The referential grounding allows us to analyze the framing of these incidents in different languages and across different texts. During the project, we will use the automatically generated data to study linguistic framing as a phenomenon, build framing resources such as lexicons and corpora. We expect to capture larger variation in framing compared to traditional approaches for building such resources. Our first data release, which contains structured data about a large number of incidents and reference texts, can be found at http://dutchframenet.nl/data-releases/.
In this article we will introduce two of the new parts of the new multi-part version of the Lexical Markup Framework (LMF) ISO standard, namely part 3 of the standard (ISO 24613-3), which deals with etymological and diachronic data, and Part 4 (ISO 24613-4), which consists of a TEI serialisation of all of the prior parts of the model. We will demonstrate the use of both standards by describing the LMF encoding of a small number of examples taken from a sample conversion of the reference Portuguese dictionary Grande Dicionário Houaiss da Língua Portuguesa, part of a broader experiment comprising the analysis of different, heterogeneously encoded, Portuguese lexical resources. We present the examples in the Unified Modelling Language (UML) and also in a couple of cases in TEI.
We describe the linking of the TUFS Basic Vocabulary Modules, created for online language learning, with the Open Multilingual Wordnet. The TUFS modules have roughly 500 lexical entries in 30 languages, each with the lemma, a link across the languages, an example sentence, usage notes and sound files. The Open Multilingual Wordnet has 34 languages (11 shared with TUFS) organized into synsets linked by semantic relations, with examples and definitions for some languages. The links can be used to (i) evaluate existing wordnets, (ii) add data to these wordnets and (iii) create new open wordnets for Khmer, Korean, Lao, Mongolian, Russian, Tagalog, Urdua nd Vietnamese
In this paper we discuss the experience of bringing together over 40 different wordnets. We introduce some extensions to the GWA wordnet LMF format proposed in Vossen et al. (2016) and look at how this new information can be displayed. Notable extensions include: confidence, corpus frequency, orthographic variants, lexicalized and non-lexicalized synsets and lemmas, new parts of speech, and more. Many of these extensions already exist in multiple wordnets – the challenge was to find a compatible representation. To this end, we introduce a new version of the Open Multilingual Wordnet (Bond and Foster, 2013), that integrates a new set of tools that tests the extensions introduced by this new format, while also ensuring the integrity of the Collaborative Interlingual Index (CILI: Bond et al., 2016), avoiding the same new concept to be introduced through multiple projects.
The paper presents the issue of collocability and collocations in Russian and gives a survey of a wide range of dictionaries both printed and online ones that describe collocations. Our project deals with building a database that will include dictionary and statistical collocations. The former can be described in various lexicographic resources whereas the latter can be extracted automatically from corpora. Dictionaries differ among themselves, the information is given in various ways, making it hard for language learners and researchers to acquire data. A number of dictionaries were analyzed and processed to retrieve verified collocations, however the overlap between the lists of collocations extracted from them is still rather small. This fact indicates there is a need to create a unified resource which takes into account collocability and more examples. The proposed resource will also be useful for linguists and for studying Russian as a foreign language. The obtained results can be important for machine learning and for other NLP tasks, for instance, automatic clustering of word combinations and disambiguation.
Diachronic lexical information is not only important in the field of historical linguistics, but is also increasingly used in NLP, most recently for machine translation of low resource languages. Therefore, there is a need for fine-grained, large-coverage and accurate etymological lexical resources. In this paper, we propose a set of guidelines to generate such resources, for each step of the life-cycle of an etymological lexicon: creation, update, evaluation, dissemination, and exploitation. To illustrate the guidelines, we introduce EtymDB 2.0, an etymological database automatically generated from the Wiktionary, which contains 1.8 million lexemes, linked by more than 700,000 fine-grained etymological relations, across 2,536 living and dead languages. We also introduce use cases for which EtymDB 2.0 could represent a key resource, such as phylogenetic tree generation, low resource machine translation or medieval languages study.
In this paper we describe our work on the development and enrichment of OFrLex, a freely available, large-coverage morphological and syntactic Old French lexicon. We rely on several heterogeneous language resources to extract structured and exploitable information. The extraction follows a semi-automatic procedure with substantial manual steps to respond to difficulties encountered while aligning lexical entries from distinct language resources. OFrLex aims at improving natural language processing tasks on Old French such as part-of-speech tagging and dependency parsing. We provide quantitative information on OFrLex and discuss its reliability. We also describe and evaluate a semi-automatic, word-embedding-based lexical enrichment process aimed at increasing the accuracy of the resource. Results of this extension technique will be manually validated in the near future, a step that will take advantage of OFrLex’s viewing, searching and editing interface, which is already accessible online.
Producing related words is a key concern in historical linguistics. Given an input word, the task is to automatically produce either its proto-word, a cognate pair or a modern word derived from it. In this paper, we apply a method for producing related words based on sequence labeling, aiming to fill in the gaps in incomplete cognate sets in Romance languages with Latin etymology (producing Romanian cognates that are missing) and to reconstruct uncertified Latin words. We further investigate an ensemble-based aggregation for combining and re-ranking the word productions of multiple languages.
Aligning senses across resources and languages is a challenging task with beneficial applications in the field of natural language processing and electronic lexicography. In this paper, we describe our efforts in manually aligning monolingual dictionaries. The alignment is carried out at sense-level for various resources in 15 languages. Moreover, senses are annotated with possible semantic relationships such as broadness, narrowness, relatedness, and equivalence. In comparison to previous datasets for this task, this dataset covers a wide range of languages and resources and focuses on the more challenging task of linking general-purpose language. We believe that our data will pave the way for further advances in alignment and evaluation of word senses by creating new solutions, particularly those notoriously requiring data such as neural networks. Our resources are publicly available at https://github.com/elexis-eu/MWSA.
Progress on deep language understanding is inhibited by the lack of a broad coverage lexicon that connects linguistic behavior to ontological concepts and axioms. We have developed COLLIE-V, a deep lexical resource for verbs, with the coverage of WordNet and syntactic and semantic details that meet or exceed existing resources. Bootstrapping from a hand-built lexicon and ontology, new ontological concepts and lexical entries, together with semantic role preferences and entailment axioms, are automatically derived by combining multiple constraints from parsing dictionary definitions and examples. We evaluated the accuracy of the technique along a number of different dimensions and were able to obtain high accuracy in deriving new concepts and lexical entries. COLLIE-V is publicly available.
We developed an extensible, comprehensive Wiktionary parser that improves over several existing parsers. We predict the etymology of a word across the full range of etymology types and languages in Wiktionary, showing improvements over a strong baseline. We also model word emergence and show the application of etymology in modeling this phenomenon. We release our parser to further research in this understudied field.
The paper presents a dataset of 11,000 Polish-English translational equivalents in the form of pairs of plWordNet and Princeton WordNet lexical units linked by three types of equivalence links: strong equivalence, regular equivalence, and weak equivalence. The resource consists of the two subsets. The first subset was built in result of manual annotation of an extended sample of Polish-English sense pairs partly randomly extracted from synsets linked by interlingual relations such as I-synononymy, I-partial synonymy and I-hyponymy and partly manually selected from the surrounding synsets in the hypernymy hierarchy. The second subset was created as a result of the manual checkup of an automatically generated lists of pairs of sense equivalents on the basis of a couple of simple, rule-based heuristics. For both subsets, the same methodology of equivalence annotation was adopted based on the verification of a set of formal, semantic-pragmatic and translational features. The constructed dataset is a novum in the wordnet domain and can facilitate the precision of bilingual NLP tasks such as automatic translation, bilingual word sense disambiguation and sentiment annotation.
Transliteration is the process of expressing a proper name from a source language in the characters of a target language (e.g. from Cyrillic to Latin characters). We present TRANSLIT, a large-scale corpus with approx. 1.6 million entries in more than 180 languages with about 3 million variations of person and geolocation names. The corpus is based on various public data sources, which have been transformed into a unified format to simplify their usage, plus a newly compiled dataset from Wikipedia. In addition, we apply several machine learning methods to establish baselines for automatically detecting transliterated names in various languages. Our best systems achieve an accuracy of 92% on identification of transliterated pairs.
In this paper, we introduce a new set of lexicons for expressing subjectivity in text documents written in Brazilian Portuguese. Besides the non-English idiom, in contrast to other subjectivity lexicons available, these lexicons represent different subjectivity dimensions (other than sentiment) and are more compact in number of terms. This last feature was designed intentionally to leverage the power of word embedding techniques, i.e., with the words mapped to an embedding space and the appropriate distance measures, we can easily capture semantically related words to the ones in the lexicons. Thus, we do not need to build comprehensive vocabularies and can focus on the most representative words for each lexicon dimension. We showcase the use of these lexicons in three highly non-trivial tasks: (1) Automated Essay Scoring in the Presence of Biased Ratings, (2) Subjectivity Bias in Brazilian Presidential Elections and (3) Fake News Classification Based on Text Subjectivity. All these tasks involve text documents written in Portuguese.
In this paper, we report the release of the ACoLi Dictionary Graph, a large-scale collection of multilingual open source dictionaries available in two machine-readable formats, a graph representation in RDF, using the OntoLex-Lemon vocabulary, and a simple tabular data format to facilitate their use in NLP tasks, such as translation inference across dictionaries. We describe the mapping and harmonization of the underlying data structures into a unified representation, its serialization in RDF and TSV, and the release of a massive and coherent amount of lexical data under open licenses.
The effort in the field of Linguistics to develop theories that aim to explain language-dependent effects on language processing is greatly facilitated by the availability of reliable resources representing different languages. This project presents a detailed description of the process of creating a large and representative corpus in Romanian – a relatively under-resourced language with unique structural and typological characteristics, that can be used as a reliable language resource for linguistic studies. The decisions that have guided the construction of the corpus, including the type of corpus, its size and component resource files are discussed. Issues related to data collection, data organization and storage, as well as characteristics of the data included in the corpus are described. Currently, the corpus has approximately 5,500,000 tokens originating from written text and 100,000 tokens of spoken language. it includes language samples that represent a wide variety of registers (i.e. written language - 16 registers and 5 registers of spoken language), as well as different authors and speakers
Although Denmark is one of the most digitized countries in Europe, no coordinated efforts have been made in recent years to support the Danish language with regard to language technology and artificial intelligence. In March 2019, however, the Danish government adopted a new, ambitious strategy for LT and artificial intelligence. In this paper, we describe the process behind the development of the language-related parts of the strategy: A Danish Language Technology Committee was constituted and a comprehensive series of workshops were organized in which users, suppliers, developers, and researchers gave their valuable input based on their experiences. We describe how, based on this experience, the focus areas and recommendations for the LT strategy were established, and which steps are currently taken in order to put the strategy into practice.
In this paper, we present a corpus for use in automatic readability assessment and automatic text simplification for German, the first of its kind for this language. The corpus is compiled from web sources and consists of parallel as well as monolingual-only (simplified German) data amounting to approximately 6,200 documents (nearly 211,000 sentences). As a unique feature, the corpus contains information on text structure (e.g., paragraphs, lines), typography (e.g., font type, font style), and images (content, position, and dimensions). While the importance of considering such information in machine learning tasks involving simplified language, such as readability assessment, has repeatedly been stressed in the literature, we provide empirical evidence for its benefit. We also demonstrate the added value of leveraging monolingual-only data for automatic text simplification via machine translation through applying back-translation, a data augmentation technique.
This paper introduces a new CLARIN Knowledge Center which is the K-Centre for Atypical Communication Expertise (ACE for short) which has been established at the Centre for Language and Speech Technology (CLST) at Radboud University. Atypical communication is an umbrella term used here to denote language use by second language learners, people with language disorders or those suffering from language disabilities, but also more broadly by bilinguals and users of sign languages. It involves multiple modalities (text, speech, sign, gesture) and encompasses different developmental stages. ACE closely collaborates with The Language Archive (TLA) at the Max Planck Institute for Psycholinguistics in order to safeguard GDPR-compliant data storage and access. We explain the mission of ACE and show its potential on a number of showcases and a use case.
Corpora of disordered speech (CDS) are costly to collect and difficult to share due to personal data protection and intellectual property (IP) issues. In this contribution we discuss the legal grounds for processing CDS in the light of the GDPR, and illustrate these with two use cases from the DELAD context. One use case deals with clinical datasets and another with legacy data from Polish hearing-impaired children. For both cases, processing based on consent and on public interest are taken into consideration.
Multilingualism is a cultural cornerstone of Europe and firmly anchored in the European treaties including full language equality. However, language barriers impacting business, cross-lingual and cross-cultural communication are still omnipresent. Language Technologies (LTs) are a powerful means to break down these barriers. While the last decade has seen various initiatives that created a multitude of approaches and technologies tailored to Europe’s specific needs, there is still an immense level of fragmentation. At the same time, AI has become an increasingly important concept in the European Information and Communication Technology area. For a few years now, AI – including many opportunities, synergies but also misconceptions – has been overshadowing every other topic. We present an overview of the European LT landscape, describing funding programmes, activities, actions and challenges in the different countries with regard to LT, including the current state of play in industry and the LT market. We present a brief overview of the main LT-related activities on the EU level in the last ten years and develop strategic guidance with regard to four key dimensions.
The identification and annotation of languages in an unambiguous and standardized way is essential for the description of linguistic data. It is the prerequisite for machine-based interpretation, aggregation, and re-use of the data with respect to different languages. This makes it a key aspect especially for Linked Data and the multilingual Semantic Web. The standard for language tags is defined by IETF’s BCP 47 and ISO 639 provides the language codes that are the tags’ main constituents. However, for the identification of lesser-known languages, endangered languages, regional varieties or historical stages of a language, the ISO 639 codes are insufficient. Also, the optional language sub-tags compliant with BCP 47 do not offer a possibility fine-grained enough to represent linguistic variation. We propose a versatile pattern that extends the BCP 47 sub-tag ‘privateuse’ and is, thus, able to overcome the limits of BCP 47 and ISO 639. Sufficient coverage of the pattern is demonstrated with the use case of linguistic Linked Data of the endangered Gascon language. We show how to use a URI shortcode for the extended sub-tag, making the length compliant with BCP 47. We achieve this with a web application and API developed to encode and decode the language tag.
We describe a new version of the Gigafida reference corpus of Slovene. In addition to updating the corpus with new material and annotating it with better tools, the focus of the upgrade was also on its transformation from a general reference corpus, which contains all language variants including non-standard language, to the corpus of standard (written) Slovene. This decision could be implemented as new corpora dedicated specifically to non-standard language emerged recently. In the new version, the whole Gigafida corpus was deduplicated for the first time, which facilitates automatic extraction of data for the purposes of compilation of new lexicographic resources such as the collocations dictionary and the thesaurus of Slovene.
The present paper outlines the projected second part of the Corpus Query Lingua Franca (CQLF) family of standards: CQLF Ontology, which is currently in the process of standardization at the International Standards Organization (ISO), in its Technical Committee 37, Subcommittee 4 (TC37SC4) and its national mirrors. The first part of the family, ISO 24623-1 (henceforth CQLF Metamodel), was successfully adopted as an international standard at the beginning of 2018. The present paper reflects the state of the CQLF Ontology at the moment of submission for the Committee Draft ballot. We provide a brief overview of the CQLF Metamodel, present the assumptions and aims of the CQLF Ontology, its basic structure, and its potential extended applications. The full ontology is expected to emerge from a community process, starting from an initial version created by the authors of the present paper.
In this paper we present a first version of a transcription portal for audio files based on automatic speech recognition (ASR) in various languages. The portal is implemented in the CLARIN resources research network and intended for use by non-technical scholars. We explain the background and interdisciplinary nature of interview data, the perks and quirks of using ASR for transcribing the audio in a research context, the dos and don’ts for optimal use of the portal, and future developments foreseen. The portal is promoted in a range of workshops, but there are a number of challenges that have to be met. These challenges concern privacy issues, ASR quality, and cost, amongst others.
This paper presents a new annotation paradigm, casual annotation, along with a proposed architecture and a reference implementation, the Ellogon Casual Annotation Tool, which implements this paradigm and architecture. The novel aspects of the proposed paradigm originate from the vision to tightly integrate annotation with the casual, everyday activities of users. Annotating in a less “controlled” environment, and removing the bottleneck of selecting content and importing it to annotation infrastructures, casual annotation provides the ability to vastly increase the content that can be annotated and ease the annotation process through automatic pre-training. The proposed paradigm, architecture and reference implementation has been evaluated for more than two years on an annotation task related to sentiment analysis. Evaluation results suggest that, at least for this annotation task, there is a huge improvement in productivity after casual annotation adoption, in comparison to the more traditional annotation paradigms followed in the early stages of the annotation task.
With 24 official EU and many additional languages, multilingualism in Europe and an inclusive Digital Single Market can only be enabled through Language Technologies (LTs). European LT business is dominated by hundreds of SMEs and a few large players. Many are world-class, with technologies that outperform the global players. However, European LT business is also fragmented – by nation states, languages, verticals and sectors, significantly holding back its impact. The European Language Grid (ELG) project addresses this fragmentation by establishing the ELG as the primary platform for LT in Europe. The ELG is a scalable cloud platform, providing, in an easy-to-integrate way, access to hundreds of commercial and non-commercial LTs for all European languages, including running tools and services as well as data sets and resources. Once fully operational, it will enable the commercial and non-commercial European LT community to deposit and upload their technologies and data sets into the ELG, to deploy them through the grid, and to connect with other resources. The ELG will boost the Multilingual Digital Single Market towards a thriving European LT community, creating new jobs and opportunities. Furthermore, the ELG project organises two open calls for up to 20 pilot projects. It also sets up 32 national competence centres and the European LT Council for outreach and coordination purposes.
This paper presents the key results of a study on the global competitiveness of the European Language Technology market for three areas – Machine Translation, speech technology, and cross-lingual search. EU competitiveness is analyzed in comparison to North America and Asia. The study focuses on seven dimensions (research, innovations, investments, market dominance, industry, infrastructure, and Open Data) that have been selected to characterize the language technology market. The study concludes that while Europe still has strong positions in Research and Innovation, it lags behind North America and Asia in scaling innovations and conquering market share.
This article describes the process of gathering and constructing a bilingual parallel corpus of Islamic Hadith, which is the set of narratives reporting different aspects of the prophet Muhammad’s life. The corpus data is gathered from the six canonical Hadith collections using a custom segmentation tool that automatically segments and annotates the two Hadith components with 92% accuracy. This Hadith segmenter minimises the costs of language resource creation and produces consistent results independently from previous knowledge and experiences that usually influence human annotators. The corpus includes more than 10M tokens and will be freely available via the LREC repository.
We introduce an array of open and accessible tools to facilitate the use of the Icelandic Gigaword Corpus, in the field of Natural Language Processing as well as for students, linguists, sociologists and others benefitting from using large corpora. A KWIC engine, powered by the Swedish Korp tool is adapted to the specifics of the corpus. An n-gram viewer, highly customizable to suit different needs, allows users to study word usage throughout the period of our text collection. A frequency dictionary provides much sought after information about word frequency statistics, computed for each subcorpus as well as aggregate, disambiguating homographs based on their respective lemmas and morphosyntactic tags. Furthermore, we provide n-grams based on the corpus, and a variety of pre-trained word embeddings models, based on word2vec, GloVe, fastText and ELMo. For three of the model types, multiple word embedding models are available trained with different algorithms and using either lemmatised or unlemmatised texts.
CLARIN is a European Research Infrastructure providing access to language resources and technologies for researchers in the humanities and social sciences. It supports the use and study of language data in general and aims to increase the potential for comparative research of cultural and societal phenomena across the boundaries of languages and disciplines, all in line with the European agenda for Open Science. Data infrastructures such as CLARIN have recently embarked on the emerging frameworks for the federation of infrastructural services, such as the European Open Science Cloud and the integration of services resulting from multidisciplinary collaboration in federated services for the wider SSH domain. In this paper we describe the interoperability requirements that arise through the existing ambitions and the emerging frameworks. The interoperability theme will be addressed at several levels, including organisation and ecosystem, design of workflow services, data curation, performance measurement and collaboration.
In this paper, we describe a new national language technology programme for Icelandic. The programme, which spans a period of five years, aims at making Icelandic usable in communication and interactions in the digital world, by developing accessible, open-source language resources and software. The research and development work within the programme is carried out by a consortium of universities, institutions, and private companies, with a strong emphasis on cooperation between academia and industries. Five core projects will be the main content of the programme: language resources, speech recognition, speech synthesis, machine translation, and spell and grammar checking. We also describe other national language technology programmes and give an overview over the history of language technology in Iceland.
Privacy by Design (also referred to as Data Protection by Design) is an approach in which solutions and mechanisms addressing privacy and data protection are embedded through the entire project lifecycle, from the early design stage, rather than just added as an additional lawyer to the final product. Formulated in the 1990 by the Privacy Commissionner of Ontario, the principle of Privacy by Design has been discussed by institutions and policymakers on both sides of the Atlantic, and mentioned already in the 1995 EU Data Protection Directive (95/46/EC). More recently, Privacy by Design was introduced as one of the requirements of the General Data Protection Regulation (GDPR), obliging data controllers to define and adopt, already at the conception phase, appropriate measures and safeguards to implement data protection principles and protect the rights of the data subject. Failing to meet this obligation may result in a hefty fine, as it was the case in the Uniontrad decision by the French Data Protection Authority (CNIL). The ambition of the proposed paper is to analyse the practical meaning of Privacy by Design in the context of Language Resources, and propose measures and safeguards that can be implemented by the community to ensure respect of this principle.
The current scientific and technological landscape is characterised by the increasing availability of data resources and processing tools and services. In this setting, metadata have emerged as a key factor facilitating management, sharing and usage of such digital assets. In this paper we present ELG-SHARE, a rich metadata schema catering for the description of Language Resources and Technologies (processing and generation services and tools, models, corpora, term lists, etc.), as well as related entities (e.g., organizations, projects, supporting documents, etc.). The schema powers the European Language Grid platform that aims to be the primary hub and marketplace for industry-relevant Language Technology in Europe. ELG-SHARE has been based on various metadata schemas, vocabularies, and ontologies, as well as related recommendations and guidelines.
Defining relations between language resources provides an archive with the ability to better serve its users. This paper covers the development and implementation of a Related Works addition to the Linguistic Data Consortium’s (LDC) catalog. The authors go step-by-step through the development of the Related Works schema, implementation of the software and database changes, and data entry of the relations. The Related Work schema involved developing of a set of controlled terms for relations based on previous work and other schema. Software and database changes consisted of both front and back end interface additions, along with modification and additions to the LDC Catalog database tables. Data entry consisted of two parts: seed data from previous work and 2019 language resources, and ongoing legacy population. Previous work in this area is discussed as well as overview information about the LDC Catalog. A list of the full LDC Related Works terms is included with brief explanations.
Data is key in training modern language technologies. In this paper, we summarise the findings of the first pan-European study on obstacles to sharing language data across 29 EU Member States and CEF-affiliated countries carried out under the ELRC White Paper action on Sustainable Language Data Sharing to Support Language Equality in Multilingual Europe. Why Language Data Matters. We present the methodology of the study, the obstacles identified and report on recommendations on how to overcome those. The obstacles are classified into (1) lack of appreciation of the value of language data, (2) structural challenges, (3) disposition towards CAT tools and lack of digital skills, (4) inadequate language data management practices, (5) limited access to outsourced translations, and (6) legal concerns. Recommendations are grouped into addressing the European/national policy level, and the organisational/institutional level.
This latest in a series of Linguistic Data Consortium (LDC) progress reports to the LREC community does not describe any single language resource, evaluation campaign or technology but sketches the activities, since the last report, of a data center devoted to supporting the work of LREC attendees among other research communities. Specifically, we describe 96 new corpora released in 2018-2020 to date, a new technology evaluation campaign, ongoing activities to support multiple common task human language technology programs, and innovations to advance the methodology of language data collection and annotation.
The major European language infrastructure initiatives like CLARIN (Hinrichs and Krauwer, 2014), DARIAH (Edmond et al., 2017) or Europeana (Europeana Foundation, 2015) have been built by focusing in the first place on institutions of larger scale, like specialized research departments and larger official units like national libraries, etc. However, besides these principal players also a large number of smaller language actors could contribute to and benefit from language infrastructures. Especially since these smaller institutions, like local libraries, archives and publishers, often collect, manage and host language resources of particular value for their geographical and cultural region, it seems highly relevant to find ways of engaging and connecting them to existing European infrastructure initiatives. In this article, we first highlight the need for reaching out to smaller local language actors and discuss challenges related to this ambition. Then we present the first step in how this objective was approached within a local language infrastructure project, namely by means of a structured documentation of the local language actors landscape in South Tyrol. We describe how the documentation efforts were structured and organized, and what tool we have set up to distribute the collected data online, by adapting existing CLARIN solutions.
This contribution describes an ongoing project of speech data collection, using the web application Samrómur which is built upon Common Voice, Mozilla Foundation’s web platform for open-source voice collection. The goal of the project is to build a large-scale speech corpus for Automatic Speech Recognition (ASR) for Icelandic. Upon completion, Samrómur will be the largest open speech corpus for Icelandic collected from the public domain. We discuss the methods used for the crowd-sourcing effort and show the importance of marketing and good media coverage when launching a crowd-sourcing campaign. Preliminary results exceed our expectations, and in one month we collected data that we had estimated would take three months to obtain. Furthermore, our initial dataset of around 45 thousand utterances has good demographic coverage, is gender-balanced and with proper age distribution. We also report on the task of validating the recordings, which we have not promoted, but have had numerous hours invested by volunteers.
This paper reports on the semi-supervised development of acoustic and language models for under-resourced, code-switched speech in five South African languages. Two approaches are considered. The first constructs four separate bilingual automatic speech recognisers (ASRs) corresponding to four different language pairs between which speakers switch frequently. The second uses a single, unified, five-lingual ASR system that represents all the languages (English, isiZulu, isiXhosa, Setswana and Sesotho). We evaluate the effectiveness of these two approaches when used to add additional data to our extremely sparse training sets. Results indicate that batch-wise semi-supervised training yields better results than a non-batch-wise approach. Furthermore, while the separate bilingual systems achieved better recognition performance than the unified system, they benefited more from pseudolabels generated by the five-lingual system than from those generated by the bilingual systems.
In recent years, fake news detection has been an emerging research area. In this paper, we put forward a novel statistical approach for the generation of feature vectors to describe a document. Our so-called class label frequency distance (clfd), is shown experimentally to provide an effective way for boosting the performance of machine learning methods. Specifically, our experiments, carried out in the fake news detection domain, verify that efficient traditional machine learning methods that use our vectorization approach, consistently outperform deep learning methods that use word embeddings for small and medium sized datasets, while the results are comparable for large datasets. In addition, we demonstrate that a novel hybrid method that utilizes both a clfd-boosted logistic regression classifier and a deep learning one, clearly outperforms deep learning methods even in large datasets.
This paper presents the first attempt at Automatic Text Simplification (ATS) for Urdu, the language of 170 million people worldwide. Being a low-resource language in terms of standard linguistic resources, recent text simplification approaches that rely on manually crafted simplified corpora or lexicons such as WordNet are not applicable to Urdu. Urdu is a morphologically rich language that requires unique considerations such as proper handling of inflectional case and honorifics. We present an unsupervised method for lexical simplification of complex Urdu text. Our method only requires plain Urdu text and makes use of word embeddings together with a set of morphological features to generate simplifications. Our system achieves a BLEU score of 80.15 and SARI score of 42.02 upon automatic evaluation on manually crafted simplified corpora. We also report results for human evaluations for correctness, grammaticality, meaning-preservation and simplicity of the output. Our code and corpus are publicly available to make our results reproducible.
In the context of multilingual language model pre-training, vocabulary size for languages with a broad set of potential characters is an unsolved problem. We propose two algorithms applicable in any unsupervised multilingual pre-training task, increasing the elasticity of budget required for building the vocabulary in Byte-Pair Encoding inspired tokenizers, significantly reducing the cost of supporting Korean in a multilingual model.
The presence of offensive language on social media platforms and the implications this poses is becoming a major concern in modern society. Given the enormous amount of content created every day, automatic methods are required to detect and deal with this type of content. Until now, most of the research has focused on solving the problem for the English language, while the problem is multilingual. We construct a Danish dataset DKhate containing user-generated comments from various social media platforms, and to our knowledge, the first of its kind, annotated for various types and target of offensive language. We develop four automatic classification systems, each designed to work for both the English and the Danish language. In the detection of offensive language in English, the best performing system achieves a macro averaged F1-score of 0.74, and the best performing system for Danish achieves a macro averaged F1-score of 0.70. In the detection of whether or not an offensive post is targeted, the best performing system for English achieves a macro averaged F1-score of 0.62, while the best performing system for Danish achieves a macro averaged F1-score of 0.73. Finally, in the detection of the target type in a targeted offensive post, the best performing system for English achieves a macro averaged F1-score of 0.56, and the best performing system for Danish achieves a macro averaged F1-score of 0.63. Our work for both the English and the Danish language captures the type and targets of offensive language, and present automatic methods for detecting different kinds of offensive language such as hate speech and cyberbullying.
Although FrameNet is recognized as one of the most fine-grained lexical databases, its coverage of lexical units is still limited. To tackle this issue, we propose a two-step frame induction process: for a set of lexical units not yet present in Berkeley FrameNet data release 1.7, first remove those that cannot fit into any existing semantic frame in FrameNet; then, assign the remaining lexical units to their correct frames. We also present the Semi-supervised Deep Embedded Clustering with Anomaly Detection (SDEC-AD) model—an algorithm that maps high-dimensional contextualized vector representations of lexical units to a low-dimensional latent space for better frame prediction and uses reconstruction error to identify lexical units that cannot evoke frames in FrameNet. SDEC-AD outperforms the state-of-the-art methods in both steps of the frame induction process. Empirical results also show that definitions provide contextual information for representing and characterizing the frame membership of lexical units.
Language identification is a well-known task for natural language documents. In this paper we explore search query language identification which is usually the first task before any other query understanding. Without loss of generalization, we run our experiments on the Adobe Stock search engine. Even though the domain is relatively generic because Adobe Stock queries cover a broad range of objects and concepts, out-of-the-box language identifiers do not perform well due to the extremely short text found in queries. Unlike other well-studied supervised approaches for this task, we examine a practical approach for the cold start problem for automatically getting large-scale query-language pairs for training. We describe the process of creating weak-labeled training data and then human-annotated evaluation data for the search query language identification task. The effectiveness of this technique is demonstrated by training a gradient boosting model for language classification given a query. We out-perform the open domain text model baselines by a large margin.
Akkadian was an East-Semitic language spoken in ancient Mesopotamia. The language is attested on hundreds of thousands of cuneiform clay tablets. Several Akkadian text corpora contain only the transliterated text. In this paper, we investigate automated phonological transcription of the transliterated corpora. The phonological transcription provides a linguistically appealing form to represent Akkadian, because the transcription is normalized according to the grammatical description of a given dialect and explicitly shows the Akkadian renderings for Sumerian logograms. Because cuneiform text does not mark the inflection for logograms, the inflected form needs to be inferred from the sentence context. To the best of our knowledge, this is the first documented attempt to automatically transcribe Akkadian. Using a context-aware neural network model, we are able to automatically transcribe syllabic tokens at near human performance with 96% recall @ 3, while the logogram transcription remains more challenging at 82% recall @ 3.
We present COSTRA 1.0, a dataset of complex sentence transformations. The dataset is intended for the study of sentence-level embeddings beyond simple word alternations or standard paraphrasing. This first version of the dataset is limited to sentences in Czech but the construction method is universal and we plan to use it also for other languages. The dataset consist of 4,262 unique sentences with average length of 10 words, illustrating 15 types of modifications such as simplification, generalization, or formal and informal language variation. The hope is that with this dataset, we should be able to test semantic properties of sentence embeddings and perhaps even to find some topologically interesting “skeleton” in the sentence embedding space. A preliminary analysis using LASER, multi-purpose multi-lingual sentence embeddings suggests that the LASER space does not exhibit the desired properties.
The automation of the diagnosis and monitoring of speech affecting diseases in real life situations, such as Depression or Parkinson’s disease, depends on the existence of rich and large datasets that resemble real life conditions, such as those collected from in-the-wild multimedia repositories like YouTube. However, the cost of manually labeling these large datasets can be prohibitive. In this work, we propose to overcome this problem by automating the annotation process, without any requirements for human intervention. We formulate the annotation problem as a Multiple Instance Learning (MIL) problem, and propose a novel solution that is based on end-to-end differentiable neural networks. Our solution has the additional advantage of generalizing the MIL framework to more scenarios where the data is stil organized in bags but does not meet the MIL bag label conditions. We demonstrate the performance of the proposed method in labeling the in-the-Wild Speech Medical (WSM) Corpus, using simple textual cues extracted from videos and their metadata. Furthermore we show what is the contribution of each type of textual cues for the final model performance, as well as study the influence of the size of the bags of instances in determining the difficulty of the learning problem
Recent advances in Optical Character Recognition (OCR) and Handwritten Text Recognition (HTR) have led to more accurate textrecognition of historical documents. The Digital Humanities heavily profit from these developments, but they still struggle whenchoosing from the plethora of OCR systems available on the one hand and when defining workflows for their projects on the other hand.In this work, we present our approach to build a ground truth for a historical German-language newspaper published in black letter. Wealso report how we used it to systematically evaluate the performance of different OCR engines. Additionally, we used this ground truthto make an informed estimate as to how much data is necessary to achieve high-quality OCR results. The outcomes of our experimentsshow that HTR architectures can successfully recognise black letter text and that a ground truth size of 50 newspaper pages suffices toachieve good OCR accuracy. Moreover, our models perform equally well on data they have not seen during training, which means thatadditional manual correction for diverging data is superfluous.
Nowadays, classical count-based word embeddings using positive pointwise mutual information (PPMI) weighted co-occurrence matrices have been widely superseded by machine-learning-based methods like word2vec and GloVe. But these methods are usually applied using very large amounts of text data. In many cases, however, there is not much text data available, for example for specific domains or low-resource languages. This paper revisits PPMI by adding Dirichlet smoothing to correct its bias towards rare words. We evaluate on standard word similarity data sets and compare to word2vec and the recent state of the art for low-resource settings: Positive and Unlabeled (PU) Learning for word embeddings. The proposed method outperforms PU-Learning for low-resource settings and obtains competitive results for Maltese and Luxembourgish.
Automatic speech processing applications often have to deal with the problem of aggregating local descriptors (i.e., representations of input speech data corresponding to specific portions across the time dimension) and turning them into a single fixed-dimension representation, known as global descriptor, on top of which downstream classification tasks can be performed. In this paper, we provide an empirical assessment of different time pooling strategies when used with state-of-the-art representation learning models. In particular, insights are provided as to when it is suitable to use simple statistics of local descriptors or when more sophisticated approaches are needed. Here, language identification is used as a case study and a database containing ten oriental languages under varying test conditions (short-duration test recordings, confusing languages, unseen languages) is used. Experiments are performed with classifiers trained on top of global descriptors to provide insights on open-set evaluation performance and show that appropriate selection of such pooling strategies yield embeddings able to outperform well-known benchmark systems as well as previously results based on attention only.
We consider the problem of disambiguating the lemma and part of speech of ambiguous words in morphologically rich languages. We propose a method for disambiguating ambiguous words in context, using a large un-annotated corpus of text, and a morphological analyser—with no manual disambiguation or data annotation. We assume that the morphological analyser produces multiple analyses for ambiguous words. The idea is to train recurrent neural networks on the output that the morphological analyser produces for unambiguous words. We present performance on POS and lemma disambiguation that reaches or surpasses the state of the art—including supervised models—using no manually annotated data. We evaluate the method on several morphologically rich languages.
Recent works on cross-lingual word embeddings have been mainly focused on linear-mapping-based approaches, where pre-trained word embeddings are mapped into a shared vector space using a linear transformation. However, there is a limitation in such approaches–they follow a key assumption: words with similar meanings share similar geometric arrangements between their monolingual word embeddings, which suggest that there is a linear relationship between languages. However, such assumption may not hold for all language pairs across all semantic concepts. We investigate whether non-linear mappings can better describe the relationship between different languages by utilising kernel Canonical Correlation Analysis (KCCA). Experimental results on five language pairs show an improvement over current state-of-art results in both supervised and self-learning scenarios, confirming that non-linear mapping is a better way to describe the relationship between languages.
We present a corpus of sentence-aligned triples of German audio, German text, and English translation, based on German audio books. The speech translation data consist of 110 hours of audio material aligned to over 50k parallel sentences. An even larger dataset comprising 547 hours of German speech aligned to German text is available for speech recognition. The audio data is read speech and thus low in disfluencies. The quality of audio and sentence alignments has been checked by a manual evaluation, showing that speech alignment quality is in general very high. The sentence alignment quality is comparable to well-used parallel translation data and can be adjusted by cutoffs on the automatic alignment score. To our knowledge, this corpus is to date the largest resource for German speech recognition and for end-to-end German-to-English speech translation.
This paper describes the acquisition, preprocessing and characteristics of SEDAR, a large scale English-French parallel corpus for the financial domain. Our extensive experiments on machine translation show that SEDAR is essential to obtain good performance on finance. We observe a large gain in the performance of machine translation systems trained on SEDAR when tested on finance, which makes SEDAR suitable to study domain adaptation for neural machine translation. The first release of the corpus comprises 8.6 million high quality sentence pairs that are publicly available for research at https://github.com/autorite/sedar-bitext.
Recent machine translation algorithms mainly rely on parallel corpora. However, since the availability of parallel corpora remains limited, only some resource-rich language pairs can benefit from them. We constructed a parallel corpus for English-Japanese, for which the amount of publicly available parallel corpora is still limited. We constructed the parallel corpus by broadly crawling the web and automatically aligning parallel sentences. Our collected corpus, called JParaCrawl, amassed over 8.7 million sentence pairs. We show how it includes a broader range of domains and how a neural machine translation model trained with it works as a good pre-trained model for fine-tuning specific domains. The pre-training and fine-tuning approaches achieved or surpassed performance comparable to model training from the initial state and reduced the training time. Additionally, we trained the model with an in-domain dataset and JParaCrawl to show how we achieved the best performance with them. JParaCrawl and the pre-trained models are freely available online for research purposes.
A large number of significant assets are available online in English, which is frequently translated into native languages to ease the information sharing among local people who are not much familiar with English. However, manual translation is a very tedious, costly, and time-taking process. To this end, machine translation is an effective approach to convert text to a different language without any human involvement. Neural machine translation (NMT) is one of the most proficient translation techniques amongst all existing machine translation systems. In this paper, we have applied NMT on two of the most morphological rich Indian languages, i.e. English-Tamil and English-Malayalam. We proposed a novel NMT model using Multihead self-attention along with pre-trained Byte-Pair-Encoded (BPE) and MultiBPE embeddings to develop an efficient translation system that overcomes the OOV (Out Of Vocabulary) problem for low resourced morphological rich Indian languages which do not have much translation available online. We also collected corpus from different sources, addressed the issues with these publicly available data and refined them for further uses. We used the BLEU score for evaluating our system performance. Experimental results and survey confirmed that our proposed translator (24.34 and 9.78 BLEU score) outperforms Google translator (9.40 and 5.94 BLEU score) respectively.
In this paper, we deal with two problems in Japanese-English machine translation of news articles. The first problem is the quality of parallel corpora. Neural machine translation (NMT) systems suffer degraded performance when trained with noisy data. Because there is no clean Japanese-English parallel data for news articles, we build a novel parallel news corpus consisting of Japanese news articles translated into English in a content-equivalent manner. This is the first content-equivalent Japanese-English news corpus translated specifically for training NMT systems. The second problem involves the domain-adaptation technique. NMT systems suffer degraded performance when trained with mixed data having different features, such as noisy data and clean data. Though the existing methods try to overcome this problem by using tags for distinguishing the differences between corpora, it is not sufficient. We thus extend a domain-adaptation method using multi-tags to train an NMT model effectively with the clean corpus and existing parallel news corpora with some types of noise. Experimental results show that our corpus increases the translation quality, and that our domain-adaptation method is more effective for learning with the multiple types of corpora than existing domain-adaptation methods are.
Machine Translation (MT) is one of the most important natural language processing applications. Independently of the applied MT approach, a MT system automatically generates an equivalent version (in some target language) of an input sentence (in some source language). Recently, a new MT approach has been proposed: neural machine translation (NMT). NMT systems have already outperformed traditional phrase-based statistical machine translation (PBSMT) systems for some pairs of languages. However, any MT approach outputs errors. In this work we present a comparative study of MT errors generated by a NMT system and a PBSMT system trained on the same English – Brazilian Portuguese parallel corpus. This is the first study of this kind involving NMT for Brazilian Portuguese. Furthermore, the analyses and conclusions presented here point out the specific problems of NMT outputs in relation to PBSMT ones and also give lots of insights into how to implement automatic post-editing for a NMT system. Finally, the corpora annotated with MT errors generated by both PBSMT and NMT systems are also available.
In natural language, we often omit some words that are easily understandable from the context. In particular, pronouns of subject, object, and possessive cases are often omitted in Japanese; these are known as zero pronouns. In translation from Japanese to other languages, we need to find a correct antecedent for each zero pronoun to generate a correct and coherent translation. However, it is difficult for conventional automatic evaluation metrics (e.g., BLEU) to focus on the success of zero pronoun resolution. Therefore, we present a hand-crafted dataset to evaluate whether translation models can resolve the zero pronoun problems in Japanese to English translations. We manually and statistically validate that our dataset can effectively evaluate the correctness of the antecedents selected in translations. Through the translation experiments using our dataset, we reveal shortcomings of an existing context-aware neural machine translation model.
We study a typical intermediary task to Machine Translation, the alignment of NPs in the bitext. After arguing that the task remains relevant even in an end-to-end paradigm, we present simple, dictionary- and word vector-based baselines and a BERT-based system. Our results make clear that even state of the art systems relying on the best end-to-end methods can be improved by bringing in old-fashioned methods such as stopword removal, lemmatization, and dictionaries
Lectures translation is a case of spoken language translation and there is a lack of publicly available parallel corpora for this purpose. To address this, we examine a framework for parallel corpus mining which is a quick and effective way to mine a parallel corpus from publicly available lectures at Coursera. Our approach determines sentence alignments, relying on machine translation and cosine similarity over continuous-space sentence representations. We also show how to use the resulting corpora in a multistage fine-tuning based domain adaptation for high-quality lectures translation. For Japanese–English lectures translation, we extracted parallel data of approximately 40,000 lines and created development and test sets through manual filtering for benchmarking translation performance. We demonstrate that the mined corpus greatly enhances the quality of translation when used in conjunction with out-of-domain parallel corpora via multistage training. This paper also suggests some guidelines to gather and clean corpora, mine parallel sentences, address noise in the mined data, and create high-quality evaluation splits. For the sake of reproducibility, we have released our code for parallel data creation.
In the context of under-resourced neural machine translation (NMT), transfer learning from an NMT model trained on a high resource language pair, or from a multilingual NMT (M-NMT) model, has been shown to boost performance to a large extent. In this paper, we focus on so-called cold start transfer learning from an M-NMT model, which means that the parent model is not trained on any of the child data. Such a set-up enables quick adaptation of M-NMT models to new languages. We investigate the effectiveness of cold start transfer learning from a many-to-many M-NMT model to an under-resourced child. We show that sufficiently large sub-word vocabularies should be used for transfer learning to be effective in such a scenario. When adopting relatively large sub-word vocabularies we observe increases in performance thanks to transfer learning from a parent M-NMT model, both when translating to and from the under-resourced language. Our proposed approach involving dynamic vocabularies is both practical and effective. We report results on two under-resourced language pairs, i.e. Icelandic-English and Irish-English.
Current approaches to machine translation (MT) either translate sentences in isolation, disregarding the context they appear in, or model context at the level of the full document, without a notion of any internal structure the document may have. In this work we consider the fact that documents are rarely homogeneous blocks of text, but rather consist of parts covering different topics. Some documents, such as biographies and encyclopedia entries, have highly predictable, regular structures in which sections are characterised by different topics. We draw inspiration from Louis and Webber (2014) who use this information to improve statistical MT and transfer their proposal into the framework of neural MT. We compare two different methods of including information about the topic of the section within which each sentence is found: one using side constraints and the other using a cache-based model. We create and release the data on which we run our experiments - parallel corpora for three language pairs (Chinese-English, French-English, Bulgarian-English) from Wikipedia biographies, which we extract automatically, preserving the boundaries of sections within the articles.
Lexical ambiguity is one of the many challenging linguistic phenomena involved in translation, i.e., translating an ambiguous word with its correct sense. In this respect, previous work has shown that the translation quality of neural machine translation systems can be improved by explicitly modeling the senses of ambiguous words. Recently, several evaluation test sets have been proposed to measure the word sense disambiguation (WSD) capability of machine translation systems. However, to date, these evaluation test sets do not include any training data that would provide a fair setup measuring the sense distributions present within the training data itself. In this paper, we present an evaluation benchmark on WSD for machine translation for 10 language pairs, comprising training data with known sense distributions. Our approach for the construction of the benchmark builds upon the wide-coverage multilingual sense inventory of BabelNet, the multilingual neural parsing pipeline TurkuNLP, and the OPUS collection of translated texts from the web. The test suite is available at http://github.com/Helsinki-NLP/MuCoW.
Background: Parallel corpora are used to train and evaluate machine translation systems. To alleviate the cost of producing parallel resources for evaluation campaigns, existing corpora are leveraged. However, little information may be available about the methods used for producing the corpus, including translation direction. Objective: To gain insight on MEDLINE parallel corpus used in the biomedical task at the Workshop on Machine Translation in 2019 (WMT 2019). Material and Methods: Contact information for the authors of MEDLINE articles included in the English/Spanish (EN/ES), English/French (EN/FR), and English/Portuguese (EN/PT) WMT 2019 test sets was obtained from PubMed and publisher websites. The authors were asked about their abstract writing practices in a survey. Results: The response rate was above 20%. Authors reported that they are mainly native speakers of languages other than English. Although manual translation, sometimes via professional translation services, was commonly used for abstract translation, authors of articles in the EN/ES and EN/PT sets also relied on post-edited machine translation. Discussion: This study provides a characterization of MEDLINE authors’ language skills and abstract writing practices. Conclusion: The information collected in this study will be used to inform test set design for the next WMT biomedical task.
Neural machine translation (NMT) needs large parallel corpora for state-of-the-art translation quality. Low-resource NMT is typically addressed by transfer learning which leverages large monolingual or parallel corpora for pre-training. Monolingual pre-training approaches such as MASS (MAsked Sequence to Sequence) are extremely effective in boosting NMT quality for languages with small parallel corpora. However, they do not account for linguistic information obtained using syntactic analyzers which is known to be invaluable for several Natural Language Processing (NLP) tasks. To this end, we propose JASS, Japanese-specific Sequence to Sequence, as a novel pre-training alternative to MASS for NMT involving Japanese as the source or target language. JASS is joint BMASS (Bunsetsu MASS) and BRSS (Bunsetsu Reordering Sequence to Sequence) pre-training which focuses on Japanese linguistic units called bunsetsus. In our experiments on ASPEC Japanese–English and News Commentary Japanese–Russian translation we show that JASS can give results that are competitive with if not better than those given by MASS. Furthermore, we show for the first time that joint MASS and JASS pre-training gives results that significantly surpass the individual methods indicating their complementary nature. We will release our code, pre-trained models and bunsetsu annotated data as resources for researchers to use in their own NLP tasks.
We introduce a machine translation dataset for three pairs of languages in the legal domain with post-edited high-quality neural machine translation and independent human references. The data was collected as part of the EU APE-QUEST project and comprises crawled content from EU websites with translation from English into three European languages: Dutch, French and Portuguese. Altogether, the data consists of around 31K tuples including a source sentence, the respective machine translation by a neural machine translation system, a post-edited version of such translation by a professional translator, and - where available - the original reference translation crawled from parallel language websites. We describe the data collection process, provide an analysis of the resulting post-edits and benchmark the data using state-of-the-art quality estimation and automatic post-editing models. One interesting by-product of our post-editing analysis suggests that neural systems built with publicly available general domain data can provide high-quality translations, even though comparison to human references suggests that this quality is quite low. This makes our dataset a suitable candidate to test evaluation metrics. The data is freely available as an ELRC-SHARE resource.
Hindi-English Machine Translation is a challenging problem, owing to multiple factors including the morphological complexity and relatively free word order of Hindi, in addition to the lack of sufficient parallel training data. Neural Machine Translation (NMT) is a rapidly advancing MT paradigm and has shown promising results for many language pairs, especially in large training data scenarios. To overcome the data sparsity issue caused by the lack of large parallel corpora for Hindi-English, we propose a method to employ additional linguistic knowledge which is encoded by different phenomena depicted by Hindi. We generalize the embedding layer of the state-of-the-art Transformer model to incorporate linguistic features like POS tag, lemma and morph features to improve the translation performance. We compare the results obtained on incorporating this knowledge with the baseline systems and demonstrate significant performance improvements. Although, the Transformer NMT models have a strong efficacy to learn language constructs, we show that the usage of specific features further help in improving the translation performance.
We made a test set for Japanese-to-English discourse translation to evaluate the power of context-aware machine translation. For each discourse phenomenon, we systematically collected examples where the translation of the second sentence depends on the first sentence. Compared with a previous study on test sets for English-to-French discourse translation (CITATION), we needed different approaches to make the data because Japanese has zero pronouns and represents different senses in different characters. We improved the translation accuracy using context-aware neural machine translation, and the improvement mainly reflects the betterment of the translation of zero pronouns.
In this work, we explore massively multilingual low-resource neural machine translation. Using translations of the Bible (which have parallel structure across languages), we train models with up to 1,107 source languages. We create various multilingual corpora, varying the number and relatedness of source languages. Using these, we investigate the best ways to use this many-way aligned resource for multilingual machine translation. Our experiments employ a grammatically and phylogenetically diverse set of source languages during testing for more representative evaluations. We find that best practices in this domain are highly language-specific: adding more languages to a training set is often better, but too many harms performance—the best number depends on the source language. Furthermore, training on related languages can improve or degrade performance, depending on the language. As there is no one-size-fits-most answer, we find that it is critical to tailor one’s approach to the source language and its typology.
In this paper, we describe the details of the Timely Disclosure Documents Corpus (TDDC). TDDC was prepared by manually aligning the sentences from past Japanese and English timely disclosure documents in PDF format published by companies listed on the Tokyo Stock Exchange. TDDC consists of approximately 1.4 million parallel sentences in Japanese and English. TDDC was used as the official dataset for the 6th Workshop on Asian Translation to encourage the development of machine translation.
Growing needs in localising audiovisual content in multiple languages through subtitles call for the development of automatic solutions for human subtitling. Neural Machine Translation (NMT) can contribute to the automatisation of subtitling, facilitating the work of human subtitlers and reducing turn-around times and related costs. NMT requires high-quality, large, task-specific training data. The existing subtitling corpora, however, are missing both alignments to the source language audio and important information about subtitle breaks. This poses a significant limitation for developing efficient automatic approaches for subtitling, since the length and form of a subtitle directly depends on the duration of the utterance. In this work, we present MuST-Cinema, a multilingual speech translation corpus built from TED subtitles. The corpus is comprised of (audio, transcription, translation) triplets. Subtitle breaks are preserved by inserting special symbols. We show that the corpus can be used to build models that efficiently segment sentences into subtitles and propose a method for annotating existing subtitling corpora with subtitle breaks, conforming to the constraint of length.
Despite increasing efforts to improve evaluation of machine translation (MT) by going beyond the sentence level to the document level, the definition of what exactly constitutes a “document level” is still not clear. This work deals with the context span necessary for a more reliable MT evaluation. We report results from a series of surveys involving three domains and 18 target languages designed to identify the necessary context span as well as issues related to it. Our findings indicate that, despite the fact that some issues and spans are strongly dependent on domain and on the target language, a number of common patterns can be observed so that general guidelines for context-aware MT evaluation can be drawn.