In this paper we present Scylla, a methodology for domain adaptation of Neural Machine Translation (NMT) systems that make use of a multilingual FrameNet enriched with qualia relations as an external knowledge base. Domain adaptation techniques used in NMT usually require fine-tuning and in-domain training data, which may pose difficulties for those working with lesser-resourced languages and may also lead to performance decay of the NMT system for out-of-domain sentences. Scylla does not require fine-tuning of the NMT model, avoiding the risk of model over-fitting and consequent decrease in performance for out-of-domain translations. Two versions of Scylla are presented: one using the source sentence as input, and another one using the target sentence. We evaluate Scylla in comparison to a state-of-the-art commercial NMT system in an experiment in which 50 sentences from the Sports domain are translated from Brazilian Portuguese to English. The two versions of Scylla significantly outperform the baseline commercial system in HTER.
Traditional automatic evaluation metrics for machine translation have been widely criticized by linguists due to their low accuracy, lack of transparency, focus on language mechanics rather than semantics, and low agreement with human quality evaluation. Human evaluations in the form of MQM-like scorecards have always been carried out in real industry setting by both clients and translation service providers (TSPs). However, traditional human translation quality evaluations are costly to perform and go into great linguistic detail, raise issues as to inter-rater reliability (IRR) and are not designed to measure quality of worse than premium quality translations. In this work, we introduce HOPE, a task-oriented and human-centric evaluation framework for machine translation output based on professional post-editing annotations. It contains only a limited number of commonly occurring error types, and uses a scoring model with geometric progression of error penalty points (EPPs) reflecting error severity level to each translation unit. The initial experimental work carried out on English-Russian language pair MT outputs on marketing content type of text from highly technical domain reveals that our evaluation framework is quite effective in reflecting the MT output quality regarding both overall system-level performance and segment-level transparency, and it increases the IRR for error type interpretation. The approach has several key advantages, such as ability to measure and compare less than perfect MT output from different systems, ability to indicate human perception of quality, immediate estimation of the labor effort required to bring MT output to premium quality, low-cost and faster application, as well as higher IRR. Our experimental data is available at https://github.com/lHan87/HOPE.
In recent years, there has been an increasing need for the restoration and translation of historical languages. In this study, we attempt to translate historical records in ancient Korean language based on neural machine translation (NMT). Inspired by priming, a cognitive science theory that two different stimuli influence each other, we propose novel priming ancient-Korean NMT (AKNMT) using bilingual subword embedding initialization with structural property awareness in the ancient documents. Finally, we obtain state-of-the-art results in the AKNMT task. To the best of our knowledge, we confirm the possibility of developing a human-centric model that incorporates the concepts of cognitive science and analyzes the result from the perspective of interference and cognitive dissonance theory for the first time.
In the present paper, we describe a large corpus of eye movement data, collected during natural reading of a human translation and a machine translation of a full novel. This data set, called GECO-MT (Ghent Eye tracking Corpus of Machine Translation) expands upon an earlier corpus called GECO (Ghent Eye-tracking Corpus) by Cop et al. (2017). The eye movement data in GECO-MT will be used in future research to investigate the effect of machine translation on the reading process and the effects of various error types on reading. In this article, we describe in detail the materials and data collection procedure of GECO-MT. Extensive information on the language proficiency of our participants is given, as well as a comparison with the participants of the original GECO. We investigate the distribution of a selection of important eye movement variables and explore the possibilities for future analyses of the data. GECO-MT is freely available at https://www.lt3.ugent.be/resources/geco-mt.
This article presents the first output of the Dutch FrameNet annotation tool, which facilitates both referential- and frame annotations of language-independent corpora. On the referential level, the tool links in-text mentions to structured data, grounding the text in the real world. On the frame level, those same mentions are annotated with respect to their semantic sense. This way of annotating not only generates a rich linguistic dataset that is grounded in real-world event instances, but also guides the annotators in frame identification, resulting in high inter-annotator-agreement and consistent annotations across documents and at discourse level, exceeding traditional sentence level annotations of frame elements. Moreover, the annotation tool features a dynamic lexical lookup that increases the development of a cross-domain FrameNet lexicon.
We present The Central Word Register for Danish (COR), which is an open source lexicon project for general AI purposes funded and initiated by the Danish Agency for Digitisation as part of an AI initiative embarked by the Danish Government in 2020. We focus here on the lexical semantic part of the project (COR-S) and describe how we – based on the existing fine-grained sense inventory from Den Danske Ordbog (DDO) – compile a more AI suitable sense granularity level of the vocabulary. A three-step methodology is applied: We establish a set of linguistic principles for defining core senses in COR-S and from there, we generate a hand-crafted gold standard of 6,000 lemmas depicting how to come from the fine-grained DDO sense to the COR inventory. Finally, we experiment with a number of language models in order to automatize the sense reduction of the rest of the lexicon. The models comprise a ruled-based model that applies our linguistic principles in terms of features, a word2vec model using cosine similarity to measure the sense proximity, and finally a deep neural BERT model fine-tuned on our annotations. The rule-based approach shows best results, in particular on adjectives, however, when focusing on the average polysemous vocabulary, the BERT model shows promising results too.
In this paper we examine existing sentiment lexicons and sense-based sentiment-tagged corpora to find out how sense and concept-based semantic relations effect sentiment scores (for polarity and valence). We show that some relations are good predictors of sentiment of related words: antonyms have similar valence and opposite polarity, synonyms similar valence and polarity, as do many derivational relations. We use this knowledge and existing resources to build a sentiment annotated wordnet of English, and show how it can be used to produce sentiment lexicons for other languages using the Open Multilingual Wordnet.
This paper reports on the most recent improvements on the Cantonese Wordnet, a wordnet project started in 2019 (Sio and Morgado da Costa, 2019) with the aim of capturing and organizing lexico-semantic information of Hong Kong Cantonese. The improvements we present here extend both the breadth and depth of the Cantonese Wordnet: increasing the general coverage, adding functional categories, enriching verbal representations, as well as creating the Cantonese Wordnet Corpus – a corpus of handcrafted examples where individual senses are shown in context.
We present ZAEBUC, an annotated Arabic-English bilingual writer corpus comprising short essays by first-year university students at Zayed University in the United Arab Emirates. We describe and discuss the various guidelines and pipeline processes we followed to create the annotations and quality check them. The annotations include spelling and grammar correction, morphological tokenization, Part-of-Speech tagging, lemmatization, and Common European Framework of Reference (CEFR) ratings. All of the annotations are done on Arabic and English texts using consistent guidelines as much as possible, with tracked alignments among the different annotations, and to the original raw texts. For morphological tokenization, POS tagging, and lemmatization, we use existing automatic annotation tools followed by manual correction. We also present various measurements and correlations with preliminary insights drawn from the data and annotations. The publicly available ZAEBUC corpus and its annotations are intended to be the stepping stones for additional annotations.
Universal Conceptual Cognitive Annotation (UCCA) (Abend and Rappoport, 2013a) is a cross-lingual semantic annotation framework that provides an easy annotation without any requirement for linguistic background. UCCA-annotated datasets have been already released in English, French, and German. In this paper, we introduce the first UCCA-annotated Turkish dataset that currently involves 50 sentences obtained from the METU-Sabanci Turkish Treebank (Atalay et al., 2003; Oflazeret al., 2003). We followed a semi-automatic annotation approach, where an external semantic parser is utilised for an initial annotation of the dataset, which is partially accurate and requires refinement. We manually revised the annotations obtained from the semantic parser that are not in line with the UCCA rules that we defined for Turkish. We used the same external semantic parser for evaluation purposes and conducted experiments with both zero-shot and few-shot learning. While the parser cannot predict remote edges in zero-shot setting, using even a small subset of training data in few-shot setting increased the overall F-1 score including the remote edges. This is the initial version of the annotated dataset and we are currently extending the dataset. We will release the current Turkish UCCA annotation guideline along with the annotated dataset.
This article presents the current outcomes of the CURLICAT CEF Telecom project, which aims to collect and deeply annotate a set of large corpora from selected domains. The CURLICAT corpus includes 7 monolingual corpora (Bulgarian, Croatian, Hungarian, Polish, Romanian, Slovak and Slovenian) containing selected samples from respective national corpora. These corpora are automatically tokenized, lemmatized and morphologically analysed and the named entities annotated. The annotations are uniformly provided for each language specific corpus while the common metadata schema is harmonised across the languages. Additionally, the corpora are annotated for IATE terms in all languages. The file format is CoNLL-U Plus format, containing the ten columns specific to the CoNLL-U format and three extra columns specific to our corpora as defined by Varádi et al. (2020). The CURLICAT corpora represent a rich and valuable source not just for training NMT models, but also for further studies and developments in machine learning, cross-lingual terminological data extraction and classification.
Social media has provided a platform for many individuals to easily express themselves naturally and publicly, and researchers have had the opportunity to utilize large quantities of this data to improve author trait analysis techniques and to improve author trait profiling systems. The majority of the work in this area, however, has been narrowly spent on English and other Western European languages, and generally focuses on a single social network at a time, despite the large quantity of data now available across languages and differences that have been found across platforms. This paper introduces RU-ADEPT, a dataset of Russian authors’ personality trait scores–Big Five and Dark Triad, demographic information (e.g. age, gender), with associated corpus of the authors’ cross-contributions to (up to) four different social media platforms–VKontakte (VK), LiveJournal, Blogger, and Moi Mir. We believe this to be the first publicly-available dataset associating demographic and personality trait data with Russian-language social media content, the first paper to describe the collection of Dark Triad scores with texts across multiple Russian-language social media platforms, and to a limited extent, the first publicly-available dataset of personality traits to author content across several different social media sites.
Questions asked by humans during a conversation often contain contextual dependencies, i.e., explicit or implicit references to previous dialogue turns. These dependencies take the form of coreferences (e.g., via pronoun use) or ellipses, and can make the understanding difficult for automated systems. One way to facilitate the understanding and subsequent treatments of a question is to rewrite it into an out-of-context form, i.e., a form that can be understood without the conversational context. We propose CoQAR, a corpus containing 4.5K conversations from the Conversational Question-Answering dataset CoQA, for a total of 53K follow-up question-answer pairs. Each original question was manually annotated with at least 2 at most 3 out-of-context rewritings. CoQA originally contains 8k conversations, which sum up to 127k question-answer pairs. CoQAR can be used in the supervised learning of three tasks: question paraphrasing, question rewriting and conversational question answering. In order to assess the quality of CoQAR’s rewritings, we conduct several experiments consisting in training and evaluating models for these three tasks. Our results support the idea that question rewriting can be used as a preprocessing step for (conversational and non-conversational) question answering models, thereby increasing their performances.
Most systems helping to provide structured information and support opinion building, discuss with users without considering their individual interest. The scarce existing research on user interest in dialogue systems depends on explicit user feedback. Such systems require user responses that are not content-related and thus, tend to disturb the dialogue flow. In this paper, we present a novel model for implicitly estimating user interest during argumentative dialogues based on semantically clustered data. Therefore, an online user study was conducted to acquire training data which was used to train a binary neural network classifier in order to predict whether or not users are still interested in the content of the ongoing dialogue. We achieved a classification accuracy of 74.9% and furthermore investigated with different Artificial Neural Networks (ANN) which new argument would fit the user interest best.
Incorporating handwritten domain scripts into neural-based task-oriented dialogue systems may be an effective way to reduce the need for large sets of annotated dialogues. In this paper, we investigate how the use of domain scripts written by conversational designers affects the performance of neural-based dialogue systems. To support this investigation, we propose the Conversational-Logic-Injection-in-Neural-Network system (CLINN) where domain scripts are coded in semi-logical rules. By using CLINN, we evaluated semi-logical rules produced by a team of differently-skilled conversational designers. We experimented with the Restaurant domain of the MultiWOZ dataset. Results show that external knowledge is extremely important for reducing the need for annotated examples for conversational systems. In fact, rules from conversational designers used in CLINN significantly outperform a state-of-the-art neural-based dialogue system when trained with smaller sets of annotated dialogues.
Open-domain dialogue systems aim to converse with humans through text, and dialogue research has heavily relied on benchmark datasets. In this work, we observe the overlapping problem in DailyDialog and OpenSubtitles, two popular open-domain dialogue benchmark datasets. Our systematic analysis then shows that such overlapping can be exploited to obtain fake state-of-the-art performance. Finally, we address this issue by cleaning these datasets and setting up a proper data processing procedure for future research.
This paper is framed in the context of the SSHOC project and aims at exploring how Language Technologies can help in promoting and facilitating multilingualism in the Social Sciences and Humanities (SSH). Although most SSH researchers produce culturally and societally relevant work in their local languages, metadata and vocabularies used in the SSH domain to describe and index research data are currently mostly in English. We thus investigate Natural Language Processing and Machine Translation approaches in view of providing resources and tools to foster multilingual access and discovery to SSH content across different languages. As case studies, we create and deliver as freely, openly available data a set of multilingual metadata concepts and an automatically extracted multilingual Data Stewardship terminology. The two case studies allow as well to evaluate performances of state-of-the-art tools and to derive a set of recommendations as to how best apply them. Although not adapted to the specific domain, the employed tools prove to be a valid asset to translation tasks. Nonetheless, validation of results by domain experts proficient in the language is an unavoidable phase of the whole workflow.
The publication of resources for minority languages requires a balance between making data open and accessible and respecting the rights and needs of its language community. The FAIR principles were introduced as a guide to good open data practices and they have since been complemented by the CARE principles for indigenous data governance. This article describes how the DGS Corpus implemented these principles and how the two sets of principles affected each other. The DGS Corpus is a large collection of recordings of members of the deaf community in Germany communicating in their primary language, German Sign Language (DGS); it was created to be both as a resource for linguistic research and as a record of the life experiences of deaf people in Germany. The corpus was designed with CARE in mind to respect and empower the language community and FAIR data publishing was used to enhance its usefulness as a scientific resource.
The European Language Grid enables researchers and practitioners to easily distribute and use NLP resources and models, such as corpora and classifiers. We describe in this paper how, during the course of our EVALITA4ELG project, we have integrated datasets and systems for the Italian language. We show how easy it is to use the integrated systems, and demonstrate in case studies how seamless the application of the platform is, providing Italian NLP for everyone.
In this paper, we provide an overview of current technologies for cross-lingual link discovery, and we discuss challenges, experiences and prospects of their application to under-resourced languages. We rst introduce the goals of cross-lingual linking and associated technologies, and in particular, the role that the Linked Data paradigm (Bizer et al., 2011) applied to language data can play in this context. We de ne under-resourced languages with a speci c focus on languages actively used on the internet, i.e., languages with a digitally versatile speaker community, but limited support in terms of language technology. We argue that languages for which considerable amounts of textual data and (at least) a bilingual word list are available, techniques for cross-lingual linking can be readily applied, and that these enable the implementation of downstream applications for under-resourced languages via the localisation and adaptation of existing technologies and resources.
This paper examines the role of emotion annotations to characterize extremist content released on social platforms. The analysis of extremist content is important to identify user emotions towards some extremist ideas and to highlight the root cause of where emotions and extremist attitudes merge together. To address these issues our methodology combines knowledge from sociological and linguistic annotations to explore French extremist content collected online. For emotion linguistic analysis, the solution presented in this paper relies on a complex linguistic annotation scheme. The scheme was used to annotate extremist text corpora in French. Data sets were collected online by following semi-automatic procedures for content selection and validation. The paper describes the integrated annotation scheme, the annotation protocol that was set-up for French corpora annotation and the results, e.g. agreement measures and remarks on annotation disagreements. The aim of this work is twofold: first, to provide a characterization of extremist contents; second, to validate the annotation scheme and to test its capacity to capture and describe various aspects of emotions.
Multiword expression (MWE) identification in tweets is a complex task due to the complex linguistic nature of MWEs combined with the non-standard language use in social networks. MWE features were shown to be helpful for hate speech detection (HSD). In this article, we present joint experiments on these two related tasks on English Twitter data: first we focus on the MWE identification task, and then we observe the influence of MWE-based features on the HSD task. For MWE identification, we compare the performance of two systems: lexicon-based and deep neural networks-based (DNN). We experimentally evaluate seven configurations of a state-of-the-art DNN system based on recurrent networks using pre-trained contextual embeddings from BERT. The DNN-based system outperforms the lexicon-based one thanks to its superior generalisation power, yielding much better recall. For the HSD task, we propose a new DNN architecture for incorporating MWE features. We confirm that MWE features are helpful for the HSD task. Moreover, the proposed DNN architecture beats previous MWE-based HSD systems by 0.4 to 1.1 F-measure points on average on four Twitter HSD corpora.
Understanding the needs and fears of citizens, especially during a pandemic such as COVID-19, is essential for any government or legislative entity. An effective COVID-19 strategy further requires that the public understand and accept the restriction plans imposed by these entities. In this paper, we explore a causal mediation scenario in which we want to emphasize the use of NLP methods in combination with methods from economics and social sciences. Based on sentiment analysis of Tweets towards the current COVID-19 situation in the UK and Sweden, we conduct several causal inference experiments and attempt to decouple the effect of government restrictions on mobility behavior from the effect that occurs due to public perception of the COVID-19 strategy in a country. To avoid biased results we control for valid country specific epidemiological and time-varying confounders. Comprehensive experiments show that not all changes in mobility are caused by countries implemented policies but also by the support of individuals in the fight against this pandemic. We find that social media texts are an important source to capture citizens’ concerns and trust in policy makers and are suitable to evaluate the success of government policies.
User-generated content is full of misspellings. Rather than being just random noise, we hypothesise that many misspellings contain hidden semantics that can be leveraged for language understanding tasks. This paper presents a fine-grained annotated corpus of misspelling in Thai, together with an analysis of misspelling intention and its possible semantics to get a better understanding of the misspelling patterns observed in the corpus. In addition, we introduce two approaches to incorporate the semantics of misspelling: Misspelling Average Embedding (MAE) and Misspelling Semantic Tokens (MST). Experiments on a sentiment analysis task confirm our overall hypothesis: additional semantics from misspelling can boost the micro F1 score up to 0.4-2%, while blindly normalising misspelling is harmful and suboptimal.
Psychiatry and people suffering from mental disorders have often been given a pejorative label that induces social rejection.Many studies have addressed discourse content about psychiatry on social media, suggesting that they convey stigmatizingrepresentations of mental health disorders. In this paper, we focus for the first time on the use of psychiatric terms in tweetsin French. We first describe the annotated dataset that we use. Then we propose several deep learning models to detectautomatically (1) the different types of use of psychiatric terms (medical use, misuse or irrelevant use), and (2) the polarityof the tweet. We show that polarity detection can be improved when done in a multitask framework in combination with typeof use detection. This confirms the observations made manually on several datasets, namely that the polarity of a tweet iscorrelated to the type of term use (misuses are mostly negative whereas medical uses are neutral). The results are interesting forboth tasks and it allows to consider the possibility for performant automatic approaches in order to conduct real-time surveyson social media, larger and less expensive than existing manual ones
During the first two years of the COVID-19 pandemic, large volumes of biomedical information concerning this new disease have been published on social media. Some of this information can pose a real danger, particularly when false information is shared, for instance recommendations how to treat diseases without professional medical advice. Therefore, automatic fact-checking resources and systems developed specifically for medical domain are crucial. While existing fact-checking resources cover COVID-19 related information in news or quantify the amount of misinformation in tweets, there is no dataset providing fact-checked COVID-19 related Twitter posts with detailed annotations for biomedical entities, relations and relevant evidence. We contribute CoVERT, a fact-checked corpus of tweets with a focus on the domain of biomedicine and COVID-19 related (mis)information. The corpus consists of 300 tweets, each annotated with named entities and relations. We employ a novel crowdsourcing methodology to annotate all tweets with fact-checking labels and supporting evidence, which crowdworkers search for online. This methodology results in substantial inter-annotator agreement. Furthermore, we use the retrieved evidence extracts as part of a fact-checking pipeline, finding that the real-world evidence is more useful than the knowledge directly available in pretrained language models.
Language models are ubiquitous in current NLP, and their multilingual capacity has recently attracted considerable attention. However, current analyses have almost exclusively focused on (multilingual variants of) standard benchmarks, and have relied on clean pre-training and task-specific corpora as multilingual signals. In this paper, we introduce XLM-T, a model to train and evaluate multilingual language models in Twitter. In this paper we provide: (1) a new strong multilingual baseline consisting of an XLM-R (Conneau et al. 2020) model pre-trained on millions of tweets in over thirty languages, alongside starter code to subsequently fine-tune on a target task; and (2) a set of unified sentiment analysis Twitter datasets in eight different languages and a XLM-T model trained on this dataset.
Natural language processing (NLP) has been shown to perform well in various tasks, such as answering questions, ascertaining natural language inference and anomaly detection. However, there are few NLP-related studies that touch upon the moral context conveyed in text. This paper studies whether state-of-the-art, pre-trained language models are capable of passing moral judgments on posts retrieved from a popular Reddit user board. Reddit is a social discussion website and forum where posts are promoted by users through a voting system. In this work, we construct a dataset that can be used for moral judgement tasks by collecting data from the AITA? (Am I the A*******?) subreddit. To model our task, we harnessed the power of pre-trained language models, including BERT, RoBERTa, RoBERTa-large, ALBERT and Longformer. We then fine-tuned these models and evaluated their ability to predict the correct verdict as judged by users for each post in the datasets. RoBERTa showed relative improvements across the three datasets, exhibiting a rate of 87% accuracy and a Matthews correlation coefficient (MCC) of 0.76, while the use of the Longformer model slightly improved the performance when used with longer sequences, achieving 87% accuracy and 0.77 MCC.
Question generation from knowledge bases (or knowledge base question generation, KBQG) is the task of generating questions from structured database information, typically in the form of triples representing facts. To handle rare entities and generalize to unseen properties, previous work on KBQG resorted to extensive, often ad-hoc pre- and post-processing of the input triple. We revisit KBQG – using pre training, a new (triple, question) dataset and taking question type into account – and show that our approach outperforms previous work both in a standard and in a zero-shot setting. We also show that the extended KBQG dataset (also helpful for knowledge base question answering) we provide allows not only for better coverage in terms of knowledge base (KB) properties but also for increased output variability in that it permits the generation of multiple questions from the same KB triple.
In recent years, Large Language Models such as GPT-3 showed remarkable capabilities in performing NLP tasks in the zero and few shot settings. On the other hand, the experiments highlighted the difficulty of GPT-3 in carrying out tasks that require a certain degree of reasoning, such as arithmetic operations. In this paper we evaluate the ability of Transformer Language Models to perform arithmetic operations following a pipeline that, before performing computations, decomposes numbers in units, tens, and so on. We denote the models fine-tuned with this pipeline with the name Calculon and we test them in the task of performing additions, subtractions and multiplications on the same test sets of GPT-3. Results show an increase of accuracy of 63% in the five-digit addition task. Moreover, we demonstrate the importance of the decomposition pipeline introduced, since fine-tuning the same Language Model without decomposing numbers results in 0% accuracy in the five-digit addition task.
Automatic dialogue summarization is a task used to succinctly summarize a dialogue transcript while correctly linking the speakers and their speech, which distinguishes this task from a conventional document summarization. To address this issue and reduce the “who said what”-related errors in a summary, we propose embedding the speaker identity information in the input embedding into the dialogue transcript encoder. Unlike the speaker embedding proposed by Gu et al. (2020), our proposal takes into account the informativeness of position embedding. By experimentally comparing several embedding methods, we confirmed that the scores of ROUGE and a human evaluation of the generated summaries were substantially increased by embedding speaker information at the less informative part of the fixed position embedding with sinusoidal functions.
We present results from a study investigating how users perceive text quality and readability in extractive and abstractive summaries. We trained two summarisation models on Swedish news data and used these to produce summaries of articles. With the produced summaries, we conducted an online survey in which the extractive summaries were compared to the abstractive summaries in terms of fluency, adequacy and simplicity. We found statistically significant differences in perceived fluency and adequacy between abstractive and extractive summaries but no statistically significant difference in simplicity. Extractive summaries were preferred in most cases, possibly due to the types of errors the summaries tend to have.
Neural text summarization has shown great potential in recent years. However, current state-of-the-art summarization models are limited by their maximum input length, posing a challenge to summarizing longer texts comprehensively. As part of a layered summarization architecture, we introduce PureText, a simple yet effective pre-processing layer that removes low- quality sentences in articles to improve existing summarization models. When evaluated on popular datasets like WikiHow and Reddit TIFU, we show up to 3.84 and 8.57 point ROUGE-1 absolute improvement on the full test set and the long article subset, respectively, for state-of-the-art summarization models such as BertSum and BART. Our approach provides downstream models with higher-quality sentences for summarization, improving overall model performance, especially on long text articles.
We introduce document retrieval and comment generation tasks for automating horizon scanning. This is an important task in the field of futurology that collects sufficient information for predicting drastic societal changes in the mid- or long-term future. The steps used are: 1) retrieving news articles that imply drastic changes, and 2) writing subjective comments on each article for others’ ease of understanding. As a first step in automating these tasks, we create a dataset that contains 2,266 manually collected news articles with comments written by experts. We analyze the collected documents and comments regarding characteristic words, the distance to general articles, and contents in the comments. Furthermore, we compare several methods for automating horizon scanning. Our experiments show that 1) manually collected articles are different from general articles regarding the words used and semantic distances, 2) the contents in the comment can be classified into several categories, and 3) a supervised model trained on our dataset achieves a better performance. The contributions are: 1) we propose document retrieval and comment generation tasks for horizon scanning, 2) create and analyze a new dataset, and 3) report the performance of several models and show that comment generation tasks are challenging.
Pre-trained language models have become crucial to achieving competitive results across many Natural Language Processing (NLP) problems. For monolingual pre-trained models in low-resource languages, the quantity has been significantly increased. However, most of them relate to the general domain, and there are limited strong baseline language models for domain-specific. We introduce ViHealthBERT, the first domain-specific pre-trained language model for Vietnamese healthcare. The performance of our model shows strong results while outperforming the general domain language models in all health-related datasets. Moreover, we also present Vietnamese datasets for the healthcare domain for two tasks are Acronym Disambiguation (AD) and Frequently Asked Questions (FAQ) Summarization. We release our ViHealthBERT to facilitate future research and downstream application for Vietnamese NLP in domain-specific. Our dataset and code are available in https://github.com/demdecuong/vihealthbert.
Graph convolutional networks (GCNs) are a powerful architecture for representation learning on documents that naturally occur as graphs, e.g., citation or social networks. However, sensitive personal information, such as documents with people’s profiles or relationships as edges, are prone to privacy leaks, as the trained model might reveal the original input. Although differential privacy (DP) offers a well-founded privacy-preserving framework, GCNs pose theoretical and practical challenges due to their training specifics. We address these challenges by adapting differentially-private gradient-based training to GCNs and conduct experiments using two optimizers on five NLP datasets in two languages. We propose a simple yet efficient method based on random graph splits that not only improves the baseline privacy bounds by a factor of 2.7 while retaining competitive F1 scores, but also provides strong privacy guarantees of epsilon = 1.0. We show that, under certain modeling choices, privacy-preserving GCNs perform up to 90% of their non-private variants, while formally guaranteeing strong privacy measures.
This paper studies solving Arabic Math Word Problems by deep learning. A Math Word Problem (MWP) is a text description of a mathematical problem that can be solved by deriving a math equation to reach the answer. Effective models have been developed for solving MWPs in English and Chinese. However, Arabic MWPs are rarely studied. This paper contributes the first large-scale dataset for Arabic MWPs, which contains 6,000 samples of primary-school math problems, written in Modern Standard Arabic (MSA). Arabic MWP solvers are then built with deep learning models and evaluated on this dataset. In addition, a transfer learning model is built to let the high-resource Chinese MWP solver promote the performance of the low-resource Arabic MWP solver. This work is the first to use deep learning methods to solve Arabic MWP and the first to use transfer learning to solve MWP across different languages. The transfer learning enhanced solver has an accuracy of 74.15%, which is 3% higher than the solver without using transfer learning. We make the dataset and solvers available in public for encouraging more research of Arabic MWPs: https://github.com/reem-codes/ArMATH
Training transformer language models requires vast amounts of text and computational resources. This drastically limits the usage of these models in niche domains for which they are not optimized, or where domain-specific training data is scarce. We focus here on the clinical domain because of its limited access to training data in common tasks, while structured ontological data is often readily available. Recent observations in model compression of transformer models show optimization potential in improving the representation capacity of attention heads. We propose KIMERA (Knowledge Injection via Mask Enforced Retraining of Attention) for detecting, retraining and instilling attention heads with complementary structured domain knowledge. Our novel multi-task training scheme effectively identifies and targets individual attention heads that are least useful for a given downstream task and optimizes their representation with information from structured data. KIMERA generalizes well, thereby building the basis for an efficient fine-tuning. KIMERA achieves significant performance boosts on seven datasets in the medical domain in Information Retrieval and Clinical Outcome Prediction settings. We apply KIMERA to BERT-base to evaluate the extent of the domain transfer and also improve on the already strong results of BioBERT in the clinical domain.
Running large-scale pre-trained language models in computationally constrained environments remains a challenging problem yet to be addressed, while transfer learning from these models has become prevalent in Natural Language Processing tasks. Several solutions, including knowledge distillation, network quantization, or network pruning have been previously proposed; however, these approaches focus mostly on the English language, thus widening the gap when considering low-resource languages. In this work, we introduce three light and fast versions of distilled BERT models for the Romanian language: Distil-BERT-base-ro, Distil-RoBERT-base, and DistilMulti-BERT-base-ro. The first two models resulted from the individual distillation of knowledge from two base versions of Romanian BERTs available in literature, while the last one was obtained by distilling their ensemble. To our knowledge, this is the first attempt to create publicly available Romanian distilled BERT models, which were thoroughly evaluated on five tasks: part-of-speech tagging, named entity recognition, sentiment analysis, semantic textual similarity, and dialect identification. Our experimental results argue that the three distilled models offer performance comparable to their teachers, while being twice as fast on a GPU and ~35% smaller. In addition, we further test the similarity between the predictions of our students versus their teachers by measuring their label and probability loyalty, together with regression loyalty - a new metric introduced in this work.
In this paper, we propose a method to generate personalized filled pauses (FPs) with group-wise prediction models. Compared with fluent text generation, disfluent text generation has not been widely explored. To generate more human-like texts, we addressed disfluent text generation. The usage of disfluency, such as FPs, rephrases, and word fragments, differs from speaker to speaker, and thus, the generation of personalized FPs is required. However, it is difficult to predict them because of the sparsity of position and the frequency difference between more and less frequently used FPs. Moreover, it is sometimes difficult to adapt FP prediction models to each speaker because of the large variation of the tendency within each speaker. To address these issues, we propose a method to build group-dependent prediction models by grouping speakers on the basis of their tendency to use FPs. This method does not require a large amount of data and time to train each speaker model. We further introduce a loss function and a word embedding model suitable for FP prediction. Our experimental results demonstrate that group-dependent models can predict FPs with higher scores than a non-personalized one and the introduced loss function and word embedding model improve the prediction performance.
In several ASR use cases, training and adaptation of domain-specific LMs can only rely on a small amount of manually verified text transcriptions and sometimes a limited amount of in-domain speech. Training of LSTM LMs in such limited data scenarios can benefit from alternate uncertain ASR hypotheses, as observed in our recent work. In this paper, we propose a method to train Transformer LMs on ASR confusion networks. We evaluate whether these self-attention based LMs are better at exploiting alternate ASR hypotheses as compared to LSTM LMs. Evaluation results show that Transformer LMs achieve 3-6% relative reduction in perplexity on the AMI scenario meetings but perform similar to LSTM LMs on the smaller Verbmobil conversational corpus. Evaluation on ASR N-best rescoring shows that LSTM and Transformer LMs trained on ASR confusion networks do not bring significant WER reductions. However, a qualitative analysis reveals that they are better at predicting less frequent words.
Keyword extraction is the task of retrieving words that are essential to the content of a given document. Researchers proposed various approaches to tackle this problem. At the top-most level, approaches are divided into ones that require training - supervised and ones that do not - unsupervised. In this study, we are interested in settings, where for a language under investigation, no training data is available. More specifically, we explore whether pretrained multilingual language models can be employed for zero-shot cross-lingual keyword extraction on low-resource languages with limited or no available labeled training data and whether they outperform state-of-the-art unsupervised keyword extractors. The comparison is conducted on six news article datasets covering two high-resource languages, English and Russian, and four low-resource languages, Croatian, Estonian, Latvian, and Slovenian. We find that the pretrained models fine-tuned on a multilingual corpus covering languages that do not appear in the test set (i.e. in a zero-shot setting), consistently outscore unsupervised models in all six languages.
Research suggests that using generic language models in specialized domains may be sub-optimal due to significant domain differences. As a result, various strategies for developing domain-specific language models have been proposed, including techniques for adapting an existing generic language model to the target domain, e.g. through various forms of vocabulary modifications and continued domain-adaptive pretraining with in-domain data. Here, an empirical investigation is carried out in which various strategies for adapting a generic language model to the clinical domain are compared to pretraining a pure clinical language model. Three clinical language models for Swedish, pretrained for up to ten epochs, are fine-tuned and evaluated on several downstream tasks in the clinical domain. A comparison of the language models’ downstream performance over the training epochs is conducted. The results show that the domain-specific language models outperform a general-domain language model; however, there is little difference in performance of the various clinical language models. However, compared to pretraining a pure clinical language model with only in-domain data, leveraging and adapting an existing general-domain language model requires fewer epochs of pretraining with in-domain data.
We present the development of a dataset for Kazakh named entity recognition. The dataset was built as there is a clear need for publicly available annotated corpora in Kazakh, as well as annotation guidelines containing straightforward—but rigorous—rules and examples. The dataset annotation, based on the IOB2 scheme, was carried out on television news text by two native Kazakh speakers under the supervision of the first author. The resulting dataset contains 112,702 sentences and 136,333 annotations for 25 entity classes. State-of-the-art machine learning models to automatise Kazakh named entity recognition were also built, with the best-performing model achieving an exact match F1-score of 97.22% on the test set. The annotated dataset, guidelines, and codes used to train the models are freely available for download under the CC BY 4.0 licence from https://github.com/IS2AI/KazNERD.
In recent years, natural language inference has been an emerging research area. In this paper, we present a novel data augmentation technique and combine it with a unique learning procedure for that task. Our so-called automatic contextual data augmentation (acda) method manages to be fully automatic, non-trivially contextual, and computationally efficient at the same time. When compared to established data augmentation methods, it is substantially more computationally efficient and requires no manual annotation by a human expert as they usually do. In order to increase its efficiency, we combine acda with two learning optimization techniques: contrastive learning and a hybrid loss function. The former maximizes the benefit of the supervisory signal generated by acda, while the latter incentivises the model to learn the nuances of the decision boundary. Our combined approach is shown experimentally to provide an effective way for mitigating spurious data correlations within a dataset, called dataset artifacts, and as a result improves performance. Specifically, our experiments verify that acda-boosted pre-trained language models that employ our learning optimization techniques, consistently outperform the respective fine-tuned baseline pre-trained language models across both benchmark datasets and adversarial examples.
Skill Classification (SC) is the task of classifying job competences from job postings. This work is the first in SC applied to Danish job vacancy data. We release the first Danish job posting dataset: *Kompetencer* (_en_: competences), annotated for nested spans of competences. To improve upon coarse-grained annotations, we make use of The European Skills, Competences, Qualifications and Occupations (ESCO; le Vrang et al., (2014)) taxonomy API to obtain fine-grained labels via distant supervision. We study two setups: The zero-shot and few-shot classification setting. We fine-tune English-based models and RemBERT (Chung et al., 2020) and compare them to in-language Danish models. Our results show RemBERT significantly outperforms all other models in both the zero-shot and the few-shot setting.
Legal texts are often difficult to interpret, and people who interpret them need to make choices about the interpretation. To improve transparency, the interpretation of a legal text can be made explicit by formalising it. However, creating formalised representations of legal texts manually is quite labour-intensive. In this paper, we describe a method to extract structured representations in the Flint language (van Doesburg and van Engers, 2019) from natural language. Automated extraction of knowledge representation not only makes the interpretation and modelling efforts more efficient, it also contributes to reducing inter-coder dependencies. The Flint language offers a formal model that enables the interpretation of legal text by describing the norms in these texts as acts, facts and duties. To extract the components of a Flint representation, we use a rule-based method and a transformer-based method. In the transformer-based method we fine-tune the last layer with annotated legal texts. The results show that the transformed-based method (80% accuracy) outperforms the rule-based method (42% accuracy) on the Dutch Aliens Act. This indicates that the transformer-based method is a promising approach of automatically extracting Flint frames.
Spelling correction utilities have become commonplace during the writing process, however, many spelling correction utilities suffer due to the size and quality of dictionaries available to aid correction. Many terms, acronyms, and morphological variations of terms are often missing, leaving potential spelling errors unidentified and potentially uncorrected. This research describes the implementation of WikiSpell, a dynamic spelling correction tool that relies on the Wikipedia dataset search API functionality as the sole source of knowledge to aid misspelled term identification and automatic replacement. Instead of a traditional matching process to select candidate replacement terms, the replacement process is treated as a natural language information retrieval process harnessing wildcard string matching and search result statistics. The aims of this research include: 1) the implementation of a spelling correction algorithm that utilizes the wildcard operators in the Wikipedia dataset search API, 2) a review of the current spell correction tools and approaches being utilized, and 3) testing and validation of the developed algorithm against the benchmark spelling correction tool, Hunspell. The key contribution of this research is a robust, dynamic information retrieval-based spelling correction algorithm that does not require prior training. Results of this research show that the proposed spelling correction algorithm, WikiSpell, achieved comparable results to an industry-standard spelling correction algorithm, Hunspell.
In this paper, we present CrudeOilNews, a corpus of English Crude Oil news for event extraction. It is the first of its kind for Commodity News and serves to contribute towards resource building for economic and financial text mining. This paper describes the data collection process, the annotation methodology, and the event typology used in producing the corpus. Firstly, a seed set of 175 news articles were manually annotated, of which a subset of 25 news was used as the adjudicated reference test set for inter-annotator and system evaluation. The inter-annotator agreement was generally substantial, and annotator performance was adequate, indicating that the annotation scheme produces consistent event annotations of high quality. Subsequently, the dataset is expanded through (1) data augmentation and (2) Human-in-the-loop active learning. The resulting corpus has 425 news articles with approximately 11k events annotated. As part of the active learning process, the corpus was used to train basic event extraction models for machine labeling; the resulting models also serve as a validation or as a pilot study demonstrating the use of the corpus in machine learning purposes. The annotated corpus is made available for academic research purpose at https://github.com/meisin/CrudeOilNews-Corpus
To cope with the COVID-19 pandemic, many jurisdictions have introduced new or altered existing legislation. Even though these new rules are often communicated to the public in news articles, it remains challenging for laypersons to learn about what is currently allowed or forbidden since news articles typically do not reference underlying laws. We investigate an automated approach to extract legal claims from news articles and to match the claims with their corresponding applicable laws. We examine the feasibility of the two tasks concerning claims about COVID-19-related laws from Berlin, Germany. For both tasks, we create and make publicly available the data sets and report the results of initial experiments. We obtain promising results with Transformer-based models that achieve 46.7 F1 for claim extraction and 91.4 F1 for law matching, albeit with some conceptual limitations. Furthermore, we discuss challenges of current machine learning approaches for legal language processing and their ability for complex legal reasoning tasks.
Argumentation mining is a growing area of research and has several interesting practical applications of mining legal arguments. Support and Attack relations are the backbone of any legal argument. However, there is no publicly available dataset of these relations in the context of legal arguments expressed in court judgements. In this paper, we focus on automatically constructing such a dataset of Support and Attack relations between sentences in a court judgment with reasonable accuracy. We propose three sets of rules based on linguistic knowledge and distant supervision to identify such relations from Indian Supreme Court judgments. The first rule set is based on multiple discourse connectors, the second rule set is based on common semantic structures between argumentative sentences in a close neighbourhood, and the third rule set uses the information about the source of the argument. We also explore a BERT-based sentence pair classification model which is trained on this dataset. We release the dataset of 20506 sentence pairs - 10746 Support (precision 77.3%) and 9760 Attack (precision 65.8%). We believe that this dataset and the ideas explored in designing the linguistic rules and will boost the argumentation mining research for legal arguments.
In this paper we present KIND, an Italian dataset for Named-entity recognition. It contains more than one million tokens with annotation covering three classes: person, location, and organization. The dataset (around 600K tokens) mostly contains manual gold annotations in three different domains (news, literature, and political discourses) and a semi-automatically annotated part. The multi-domain feature is the main strength of the present work, offering a resource which covers different styles and language uses, as well as the largest Italian NER dataset with manual gold annotations. It represents an important resource for the training of NER systems in Italian. Texts and annotations are freely downloadable from the Github repository.
Question answering (QA) is one of the most common NLP tasks that relates to named entity recognition, fact extraction, semantic search and some other fields. In industry, it is much valued in chat-bots and corporate information systems. It is also a challenging task that attracted the attention of a very general audience at the quiz show Jeopardy! In this article we describe a Jeopardy!-like Russian QA data set collected from the official Russian quiz database Ch-g-k. The data set includes 379,284 quiz-like questions with 29,375 from the Russian analogue of Jeopardy! (Own Game). We observe its linguistic features and the related QA-task. We conclude about perspectives of a QA challenge based on the collected data set.
In this paper, we present a new corpus of clickbait articles annotated by university students along with a corresponding shared task: clickbait articles use a headline or teaser that hides information from the reader to make them curious to open the article. We therefore propose to construct approaches that can automatically extract the relevant information from such an article, which we call clickbait resolving. We show why solving this task might be relevant for end users, and why clickbait can probably not be defeated with clickbait detection alone. Additionally, we argue that this task, although similar to question answering and some automatic summarization approaches, needs to be tackled with specialized models. We analyze the performance of some basic approaches on this task and show that models fine-tuned on our data can outperform general question answering models, while providing a systematic approach to evaluate the results. We hope that the data set and the task will help in giving users tools to counter clickbait in the future.
We present VALET, a framework for rule-based information extraction written in Python. VALET departs from legacy approaches predicated on cascading finite-state transducers, instead offering direct support for mixing heterogeneous information–lexical, orthographic, syntactic, corpus-analytic–in a succinct syntax that supports context-free idioms. We show how a handful of rules suffices to implement sophisticated matching, and describe a user interface that facilitates exploration for development and maintenance of rule sets. Arguing that rule-based information extraction is an important methodology early in the development cycle, we describe an experiment in which a VALET model is used to annotate examples for a machine learning extraction model. While learning to emulate the extraction rules, the resulting model generalizes them, recognizing valid extraction targets the rules failed to detect.
Proper recognition and interpretation of negation signals in text or communication is crucial for any form of full natural language understanding. It is also essential for computational approaches to natural language processing. In this study we focus on negation detection in Dutch spoken human-computer conversations. Since there exists no Dutch (dialogue) corpus annotated for negation we have annotated a Dutch corpus sample to evaluate our method for automatic negation detection. We use transfer learning and trained NegBERT (an existing BERT implementation used for negation detection) on English data with multilingual BERT to detect negation in Dutch dialogues. Our results show that adding in-domain training material improves the results. We show that we can detect both negation cues and scope in Dutch dialogues with high precision and recall. We provide a detailed error analysis and discuss the effects of cross-lingual and cross-domain transfer learning on automatic negation detection.
The Linguistic Data Consortium was founded in 1992 to solve the problem that limitations in access to shareable data was impeding progress in Human Language Technology research and development. At the time, DARPA had adopted the common task research management paradigm to impose additional rigor on their programs by also providing shared objectives, data and evaluation methods. Early successes underscored the promise of this paradigm but also the need for a standing infrastructure to host and distribute the shared data. During LDC’s initial five year grant, it became clear that the demand for linguistic data could not easily be met by the existing providers and that a dedicated data center could add capacity first for data collection and shortly thereafter for annotation. The expanding purview required expansions of LDC’s technical infrastructure including systems support and software development. An open question for the center would be its role in other kinds of research beyond data development. Over its 30 years history, LDC has performed multiple roles ranging from neutral, independent data provider to multisite programs, to creator of exploratory data in tight collaboration with system developers, to research group focused on data intensive investigations.
This article highlights ELRA’s latest achievements in the field of Language Resources (LRs) identification, sharing and production. It also reports on ELRA’s involvement in several national and international projects, as well as in the organization of events for the support of LRs and related Language Technologies, including for under-resourced languages. Over the past few years, ELRA, together with its operational agency ELDA, has continued to increase its catalogue offer of LRs, establishing worldwide partnerships for the production of various types of LRs (SMS, tweets, crawled data, MT aligned data, speech LRs, sentiment-based data, etc.). Through their consistent involvement in EU-funded projects, ELRA and ELDA have contributed to improve the access to multilingual information in the context of the pandemic, develop tools for the de-identification of texts in the legal and medical domains, support the EU eTranslation Machine Translation system, and set up a European platform providing access to both resources and services. In December 2019, ELRA co-organized the LT4All conference, whose main topics were Language Technologies for enabling linguistic diversity and multilingualism worldwide. Moreover, although LREC was cancelled in 2020, ELRA published the LREC 2020 proceedings for the Main conference and Workshops papers, and carried on its dissemination activities while targeting the new LREC edition for 2022.
Ethical issues in Language Resources and Language Technology are often invoked, but rarely discussed. This is at least partly because little work has been done to systematize ethical issues and principles applicable in the fields of Language Resources and Language Technology. This paper provides an overview of ethical issues that arise at different stages of Language Resources and Language Technology development, from the conception phase through the construction phase to the use phase. Based on this overview, the authors propose a tentative taxonomy of ethical issues in Language Resources and Language Technology, built around five principles: Privacy, Property, Equality, Transparency and Freedom. The authors hope that this tentative taxonomy will facilitate ethical assessment of projects in the field of Language Resources and Language Technology, and structure the discussion on ethical issues in this domain, which may eventually lead to the adoption of a universally accepted Code of Ethics of the Language Resources and Language Technology community.
This article studies the application of the #BenderRule in Natural Language Processing (NLP) articles according to two dimensions. Firstly, in a contrastive manner, by considering two major international conferences, LREC and ACL, and secondly, in a diachronic manner, by inspecting nearly 14,000 articles over a period of time ranging from 2000 to 2020 for LREC and from 1979 to 2020 for ACL. For this purpose, we created a corpus from LREC and ACL articles from the above-mentioned periods, from which we manually annotated nearly 1,000. We then developed two classifiers to automatically annotate the rest of the corpus. Our results show that LREC articles tend to respect the #BenderRule (80 to 90% of them respect it), whereas 30 to 40% of ACL articles do not. Interestingly, over the considered periods, the results appear to be stable for the two conferences, even though a rebound in ACL 2020 could be a sign of the influence of the blog post about the #BenderRule.
While aspect-based sentiment analysis of user-generated content has received a lot of attention in the past years, emotion detection at the aspect level has been relatively unexplored. Moreover, given the rise of more visual content on social media platforms, we want to meet the ever-growing share of multimodal content. In this paper, we present a multimodal dataset for Aspect-Based Emotion Analysis (ABEA). Additionally, we take the first steps in investigating the utility of multimodal coreference resolution in an ABEA framework. The presented dataset consists of 4,900 comments on 175 images and is annotated with aspect and emotion categories and the emotional dimensions of valence and arousal. Our preliminary experiments suggest that ABEA does not benefit from multimodal coreference resolution, and that aspect and emotion classification only requires textual information. However, when more specific information about the aspects is desired, image recognition could be essential.
Sentiment analysis is one of the most widely studied tasks in natural language processing. While BERT-based models have achieved state-of-the-art results in this task, little attention has been given to its performance variability across class labels, multi-source and multi-domain corpora. In this paper, we present an improved state-of-the-art and comparatively evaluate BERT-based models for sentiment analysis on Italian corpora. The proposed model is evaluated over eight sentiment analysis corpora from different domains (social media, finance, e-commerce, health, travel) and sources (Twitter, YouTube, Facebook, Amazon, Tripadvisor, Opera and Personal Healthcare Agent) on the prediction of positive, negative and neutral classes. Our findings suggest that BERT-based models are confident in predicting positive and negative examples but not as much with neutral examples. We release the sentiment analysis model as well as a newly financial domain sentiment corpus.
Sentiment analysis is one of the most widely studied applications in NLP, but most work focuses on languages with large amounts of data. We introduce the first large-scale human-annotated Twitter sentiment dataset for the four most widely spoken languages in Nigeria—Hausa, Igbo, Nigerian-Pidgin, and Yorùbá—consisting of around 30,000 annotated tweets per language, including a significant fraction of code-mixed tweets. We propose text collection, filtering, processing and labeling methods that enable us to create datasets for these low-resource languages. We evaluate a range of pre-trained models and transfer strategies on the dataset. We find that language-specific models and language-adaptive fine-tuning generally perform best. We release the datasets, trained models, sentiment lexicons, and code to incentivize research on sentiment analysis in under-represented languages.
This paper presents a scheme for emotion annotation and its manual application on a genre-diverse corpus of texts written in French. The methodology introduced here emphasizes the necessity of clarifying the main concepts implied by the analysis of emotions as they are expressed in texts, before conducting a manual annotation campaign. After explaining whatentails a deeply linguistic perspective on emotion expression modeling, we present a few NLP works that share some common points with this perspective and meticulously compare our approach with them. We then highlight some interesting quantitative results observed on our annotated corpus. The most notable interactions are on the one hand between emotion expression modes and genres of texts, and on the other hand between emotion expression modes and emotional categories. These observation corroborate and clarify some of the results already mentioned in other NLP works on emotion annotation.
In this paper we address the question of how to integrate grammar and lexical-semantic knowledge within a single and homogeneous knowledge graph. We introduce a graph modelling of grammar knowledge which enables its merging with a lexical-semantic network. Such an integrated representation is expected, for instance, to provide new material for language-related graph embeddings in order to model interactions between Syntax and Semantics. Our base model relies on a phrase structure grammar. The phrase structure is accounted for by both a Proof-Theoretical representation, through a Context-Free Grammar, and a Model-Theoretical one, through a constraint-based grammar. The constraint types colour the grammar layer with syntactic relationships such as Immediate Dominance, Linear Precedence, and more. We detail a creation process which infers the grammar layer from a corpus annotated in constituency and integrates it with a lexical-semantic network through a shared POS tagset. We implement the process, and experiment with the French Treebank and the JeuxDeMots lexical-semantic network. The outcome is the HOLINET knowledge graph.
State-of-the-art approaches for metaphor detection compare their literal - or core - meaning and their contextual meaning using metaphor classifiers based on neural networks. However, metaphorical expressions evolve over time due to various reasons, such as cultural and societal impact. Metaphorical expressions are known to co-evolve with language and literal word meanings, and even drive, to some extent, this evolution. This poses the question of whether different, possibly time-specific, representations of literal meanings may impact the metaphor detection task. To the best of our knowledge, this is the first study that examines the metaphor detection task with a detailed exploratory analysis where different temporal and static word embeddings are used to account for different representations of literal meanings. Our experimental analysis is based on three popular benchmarks used for metaphor detection and word embeddings extracted from different corpora and temporally aligned using different state-of-the-art approaches. The results suggest that the usage of different static word embedding methods does impact the metaphor detection task and some temporal word embeddings slightly outperform static methods. However, the results also suggest that temporal word embeddings may provide representations of the core meaning of the metaphor even too close to their contextual meaning, thus confusing the classifier. Overall, the interaction between temporal language evolution and metaphor detection appears tiny in the benchmark datasets used in our experiments. This suggests that future work for the computational analysis of this important linguistic phenomenon should first start by creating a new dataset where this interaction is better represented.
Task embeddings are low-dimensional representations that are trained to capture task properties. In this paper, we propose MetaEval, a collection of 101 NLP tasks. We fit a single transformer to all MetaEval tasks jointly while conditioning it on learned embeddings. The resulting task embeddings enable a novel analysis of the space of tasks. We then show that task aspects can be mapped to task embeddings for new tasks without using any annotated examples. Predicted embeddings can modulate the encoder for zero-shot inference and outperform a zero-shot baseline on GLUE tasks. The provided multitask setup can function as a benchmark for future transfer learning research.
Automatic Term Extraction (ATE) is a key component for domain knowledge understanding and an important basis for further natural language processing applications. Even with persistent improvements, ATE still exhibits weak results exacerbated by small training data inherent to specialized domain corpora. Recently, transformers-based deep neural models, such as BERT, have proven to be efficient in many downstream NLP tasks. However, no systematic evaluation of ATE has been conducted so far. In this paper, we run an extensive study on fine-tuning pre-trained BERT models for ATE. We propose strategies that empirically show BERT’s effectiveness using cross-lingual and cross-domain transfer learning to extract single and multi-word terms. Experiments have been conducted on four specialized domains in three languages. The obtained results suggest that BERT can capture cross-domain and cross-lingual terminologically-marked contexts shared by terms, opening a new design-pattern for ATE.
We approach aspect-based argument mining as a supervised machine learning task to classify arguments into semantically coherent groups referring to the same defined aspect categories. As an exemplary use case, we introduce the Argument Aspect Corpus - Nuclear Energy that separates arguments about the topic of nuclear energy into nine major aspects. Since the collection of training data for further aspects and topics is costly, we investigate the potential for current transformer-based few-shot learning approaches to accurately classify argument aspects. The best approach is applied to a British newspaper corpus covering the debate on nuclear energy over the past 21 years. Our evaluation shows that a stable prediction of shares of argument aspects in this debate is feasible with 50 to 100 training samples per aspect. Moreover, we see signals for a clear shift in the public discourse in favor of nuclear energy in recent years. This revelation of changing patterns of pro and contra arguments related to certain aspects over time demonstrates the potential of supervised argument aspect detection for tracking issue-specific media discourses.
Vocabulary learning is vital to foreign language learning. Correct and adequate feedback is essential to successful and satisfying vocabulary training. However, many vocabulary and language evaluation systems perform on simple rules and do not account for real-life user learning data. This work introduces Multi-Language Vocabulary Evaluation Data Set (MuLVE), a data set consisting of vocabulary cards and real-life user answers, labeled indicating whether the user answer is correct or incorrect. The data source is user learning data from the Phase6 vocabulary trainer. The data set contains vocabulary questions in German and English, Spanish, and French as target language and is available in four different variations regarding pre-processing and deduplication. We experiment to fine-tune pre-trained BERT language models on the downstream task of vocabulary evaluation with the proposed MuLVE data set. The results provide outstanding results of > 95.5 accuracy and F2-score. The data set is available on the European Language Grid.
The detection and extraction of abbreviations from unstructured texts can help to improve the performance of Natural Language Processing tasks, such as machine translation and information retrieval. However, in terms of publicly available datasets, there is not enough data for training deep-neural-networks-based models to the point of generalising well over data. This paper presents PLOD, a large-scale dataset for abbreviation detection and extraction that contains 160k+ segments automatically annotated with abbreviations and their long forms. We performed manual validation over a set of instances and a complete automatic validation for this dataset. We then used it to generate several baseline models for detecting abbreviations and long forms. The best models achieved an F1-score of 0.92 for abbreviations and 0.89 for detecting their corresponding long forms. We release this dataset along with our code and all the models publicly at https://github.com/surrey-nlp/PLOD-AbbreviationDetection
We present a fairly large, Potential Idiomatic Expression (PIE) dataset for Natural Language Processing (NLP) in English. The challenges with NLP systems with regards to tasks such as Machine Translation (MT), word sense disambiguation (WSD) and information retrieval make it imperative to have a labelled idioms dataset with classes such as it is in this work. To the best of the authors’ knowledge, this is the first idioms corpus with classes of idioms beyond the literal and the general idioms classification. In particular, the following classes are labelled in the dataset: metaphor, simile, euphemism, parallelism, personification, oxymoron, paradox, hyperbole, irony and literal. We obtain an overall inter-annotator agreement (IAA) score, between two independent annotators, of 88.89%. Many past efforts have been limited in the corpus size and classes of samples but this dataset contains over 20,100 samples with almost 1,200 cases of idioms (with their meanings) from 10 classes (or senses). The corpus may also be extended by researchers to meet specific needs. The corpus has part of speech (PoS) tagging from the NLTK library. Classification experiments performed on the corpus to obtain a baseline and comparison among three common models, including the BERT model, give good results. We also make publicly available the corpus and the relevant codes for working with it for NLP tasks.
Spellchecking text written by language learners is especially challenging because errors made by learners differ both quantitatively and qualitatively from errors made by already proficient learners. We introduce LeSpell, a multi-lingual (English, German, Italian, and Czech) evaluation data set of spelling mistakes in context that we compiled from seven underlying learner corpora. Our experiments show that existing spellcheckers do not work well with learner data. Thus, we introduce a highly customizable spellchecking component for the DKPro architecture, which improves performance in many settings.
For different reasons, text can be difficult to read and understand for many people, especially if the text’s language is too complex. In order to provide suitable text for the target audience, it is necessary to measure its complexity. In this paper we describe subjective experiments to assess the readability of German text. We compile a new corpus of sentences provided by a German IT service provider. The sentences are annotated with the subjective complexity ratings by two groups of participants, namely experts and non-experts for that text domain. We then extract an extensive set of linguistically motivated features that are supposedly interacting with complexity perception. We show that a linear regression model with a subset of these features can be a very good predictor of text complexity.
In this paper, we address two problems in indexing and querying spoken language corpora with overlapping speaker contributions. First, we look into how token distance and token precedence can be measured when multiple primary data streams are available and when transcriptions happen to be tokenized, but are not synchronized with the sound at the level of individual tokens. We propose and experiment with a speaker-based search mode that enables any speaker’s transcription tier to be the basic tokenization layer whereby the contributions of other speakers are mapped to this given tier. Secondly, we address two distinct methods of how speaker overlaps can be captured in the TEI-based ISO Standard for Spoken Language Transcriptions (ISO 24624:2016) and how they can be queried by MTAS – an open source Lucene-based search engine for querying text with multilevel annotations. We illustrate the problems, introduce possible solutions and discuss their benefits and drawbacks.
This paper introduces DiaBiz, a large, annotated, multimodal corpus of Polish telephone conversations conducted in varied business settings, comprising 4036 call centre interactions from nine different domains, i.e. banking, energy services, telecommunications, insurance, medical care, debt collection, tourism, retail and car rental. The corpus was developed to boost the development of third-party speech recognition engines, dialog systems and conversational intelligence tools for Polish. Its current size amounts to nearly 410 hours of recordings and over 3 million words of transcribed speech. We present the structure of the corpus, data collection and transcription procedures, challenges of punctuating and truecasing speech transcripts, dialog structure annotation and discuss some of the ecological validity considerations involved in the development of such resources.
This paper presents the Latvian Language Learner Corpus (LaVA) developed at the Institute of Mathematics and Computer Science, University of Latvia. LaVA corpus contains 1015 essays (190k tokens and 790k characters excluding whitespaces) from foreigners studying at Latvian higher education institutions and who are learning Latvian as a foreign language in the first or second semester, reaching the A1 (possibly A2) Latvian language proficiency level. The corpus has morphological and error annotations. Error analysis and the statistics of the LaVA corpus are also provided in the paper. The corpus is publicly available at: http://www.korpuss.lv/id/LaVA.
We present the EuroPat corpus of patent-specific parallel data for 6 official European languages paired with English: German, Spanish, French, Croatian, Norwegian, and Polish. The filtered parallel corpora range in size from 51 million sentences (Spanish-English) to 154k sentences (Croatian-English), with the unfiltered (raw) corpora being up to 2 times larger. Access to clean, high quality, parallel data in technical domains such as science, engineering, and medicine is needed for training neural machine translation systems for tasks like online dispute resolution and eProcurement. Our evaluation found that the addition of EuroPat data to a generic baseline improved the performance of machine translation systems on in-domain test data in German, Spanish, French, and Polish; and in translating patent data from Croatian to English. The corpus has been released under Creative Commons Zero, and is expected to be widely useful for training high-quality machine translation systems, and particularly for those targeting technical documents such as patents and contracts.
The exploding amount of user-generated content has spurred NLP research to deal with documents from various digital communication formats (tweets, chats, emails, etc.). Using these texts as language resources implies complying with legal data privacy regulations. To protect the personal data of individuals and preclude their identification, we employ pseudonymization. More precisely, we identify those text spans that carry information revealing an individual’s identity (e.g., names of persons, locations, phone numbers, or dates) and subsequently substitute them with synthetically generated surrogates. Based on CodE Alltag, a German-language email corpus, we address two tasks. The first task is to evaluate various architectures for the automatic recognition of privacy-sensitive entities in raw data. The second task examines the applicability of pseudonymized data as training data for such systems since models learned on original data cannot be published for reasons of privacy protection. As outputs of both tasks, we, first, generate a new pseudonymized version of CodE Alltag compliant with the legal requirements of the General Data Protection Regulation (GDPR). Second, we make accessible a tagger for recognizing privacy-sensitive information in German emails and similar text genres, which is trained on already pseudonymized data.
The growth of social media has brought with it a massive channel for spreading and reinforcing stereotypes. This issue becomes critical when the affected targets are minority groups such as women, the LGBT+ community and immigrants. Although from the perspective of computational linguistics, the detection of this kind of stereotypes is steadily improving, most stereotypes are expressed implicitly and identifying them automatically remains a challenge. One of the problems we found for tackling this issue is the lack of an operationalised definition of implicit stereotypes that would allow us to annotate consistently new corpora by characterising the different forms in which stereotypes appear. In this paper, we present thirteen criteria for annotating implicitness which were elaborated to facilitate the subjective task of identifying the presence of stereotypes. We also present NewsCom-Implicitness, a corpus of 1,911 sentences, of which 426 comprise explicit and implicit racial stereotypes. An experiment was carried out to evaluate the applicability of these criteria. The results indicate that different criteria obtain different inter-annotator agreement values and that there is a greater agreement when more criteria can be identified in one sentence.
Current state of the art acoustic models can easily comprise more than 100 million parameters. This growing complexity demands larger training datasets to maintain a decent generalization of the final decision function. An ideal dataset is not necessarily large in size, but large with respect to the amount of unique speakers, utilized hardware and varying recording conditions. This enables a machine learning model to explore as much of the domain-specific input space as possible during parameter estimation. This work introduces Common Phone, a gender-balanced, multilingual corpus recorded from more than 76.000 contributors via Mozilla’s Common Voice project. It comprises around 116 hours of speech enriched with automatically generated phonetic segmentation. A Wav2Vec 2.0 acoustic model was trained with the Common Phone to perform phonetic symbol recognition and validate the quality of the generated phonetic annotation. The architecture achieved a PER of 18.1 % on the entire test set, computed with all 101 unique phonetic symbols, showing slight differences between the individual languages. We conclude that Common Phone provides sufficient variability and reliable phonetic annotation to help bridging the gap between research and application of acoustic models.
This paper presents two-fold contributions: a full revision of the Palestinian morphologically annotated corpus (Curras), and a newly annotated Lebanese corpus (Baladi). Both corpora can be used as a more general Levantine corpus. Baladi consists of around 9.6K morphologically annotated tokens. Each token was manually annotated with several morphological features and using LDC’s SAMA lemmas and tags. The inter-annotator evaluation on most features illustrates 78.5% Kappa and 90.1% F1-Score. Curras was revised by refining all annotations for accuracy, normalization and unification of POS tags, and linking with SAMA lemmas. This revision was also important to ensure that both corpora are compatible and can help to bridge the nuanced linguistic gaps that exist between the two highly mutually intelligible dialects. Both corpora are publicly available through a web portal.
This paper describes a comprehensive annotation study on Japanese judgment documents in civil cases. We aim to build an annotated corpus designed for Legal Judgment Prediction (LJP), especially for torts. Our annotation scheme contains annotations of whether tort is accepted by judges as well as its corresponding rationales for explainability purpose. Our annotation scheme extracts decisions and rationales at character-level. Moreover, the scheme can capture the explicit causal relation between judge’s decisions and their corresponding rationales, allowing multiple decisions in a document. To obtain high-quality annotation, we developed an annotation scheme with legal experts, and confirmed its reliability by agreement studies with Krippendorff’s alpha metric. The result of the annotation study suggests the proposed annotation scheme can produce a dataset of Japanese LJP at reasonable reliability.
Even though hate speech (HS) online has been an important object of research in the last decade, most HS-related corpora over-simplify the phenomenon of hate by attempting to label user comments as “hate” or “neutral”. This ignores the complex and subjective nature of HS, which limits the real-life applicability of classifiers trained on these corpora. In this study, we present the M-Phasis corpus, a corpus of ~9k German and French user comments collected from migration-related news articles. It goes beyond the “hate”-“neutral” dichotomy and is instead annotated with 23 features, which in combination become descriptors of various types of speech, ranging from critical comments to implicit and explicit expressions of hate. The annotations are performed by 4 native speakers per language and achieve high (0.77 <= k <= 1) inter-annotator agreements. Besides describing the corpus creation and presenting insights from a content, error and domain analysis, we explore its data characteristics by training several classification baselines.
In this paper, we describe ParCorFull2.0, a parallel corpus annotated with full coreference chains for multiple languages, which is an extension of the existing corpus ParCorFull (Lapshinova-Koltunski et al., 2018). Similar to the previous version, this corpus has been created to address translation of coreference across languages, a phenomenon still challenging for machine translation (MT) and other multilingual natural language processing (NLP) applications. The current version of the corpus that we present here contains not only parallel texts for the language pair English-German, but also for English-French and English-Portuguese, which are all major European languages. The new language pairs belong to the Romance languages. The addition of a new language group creates a need of extension not only in terms of texts added, but also in terms of the annotation guidelines. Both French and Portuguese contain structures not found in English and German. Moreover, Portuguese is a pro-drop language bringing even more systemic differences in the realisation of coreference into our cross-lingual resources. These differences cause problems for multilingual coreference resolution and machine translation. Our parallel corpus with full annotation of coreference will be a valuable resource with a variety of uses not only for NLP applications, but also for contrastive linguists and researchers in translation studies.
We presentDialogues in Games(DinG), a corpus of manual transcriptions of real-life, oral, spontaneous multi-party dialogues between French-speaking players of the board game Catan. Our objective is to make available a quality resource for French, composed of long dialogues, to facilitate their study in the style of (Asher et al., 2016). In a general dialogue setting, participants share personal information, which makes it impossible to disseminate the resource freely and openly. In DinG, the attention of the participants is focused on the game, which prevents them from talking about themselves. In addition, we are conducting a study on the nature of the questions in dialogue, through annotation (Cruz Blandon et al., 2019), in order to develop more natural automatic dialogue systems
This paper describes the experiments carried out during the development of the latest version of Bicleaner, named Bicleaner AI, a tool that aims at detecting noisy sentences in parallel corpora. The tool, which now implements a new neural classifier, uses state-of-the-art techniques based on pre-trained transformer-based language models fine-tuned on a binary classification task. After that, parallel corpus filtering is performed, discarding the sentences that have lower probability of being mutual translations. Our experiments, based on the training of neural machine translation (NMT) with corpora filtered using Bicleaner AI for two different scenarios, show significant improvements in translation quality compared to the previous version of the tool which implemented a classifier based on Extremely Randomized Trees.
We present ReLCo— the Revita Learner Corpus—a new semi-automatically annotated learner corpus for Russian. The corpus was collected while several thousand L2 learners were performing exercises using the Revita language-learning system. All errors were detected automatically by the system and annotated by type. Part of the corpus was annotated manually—this part was created for further experiments on automatic assessment of grammatical correctness. The Learner Corpus provides valuable data for studying patterns of grammatical errors, experimenting with grammatical error detection and grammatical error correction, and developing new exercises for language learners. Automating the collection and annotation makes the process of building the learner corpus much cheaper and faster, in contrast to the traditional approach of building learner corpora. We make the data publicly available.
The Universal Morphology (UniMorph) project is a collaborative effort providing broad-coverage instantiated normalized morphological inflection tables for hundreds of diverse world languages. The project comprises two major thrusts: a language-independent feature schema for rich morphological annotation, and a type-level resource of annotated data in diverse languages realizing that schema. This paper presents the expansions and improvements on several fronts that were made in the last couple of years (since McCarthy et al. (2020)). Collaborative efforts by numerous linguists have added 66 new languages, including 24 endangered languages. We have implemented several improvements to the extraction pipeline to tackle some issues, e.g., missing gender and macrons information. We have amended the schema to use a hierarchical structure that is needed for morphological phenomena like multiple-argument agreement and case stacking, while adding some missing morphological features to make the schema more inclusive.In light of the last UniMorph release, we also augmented the database with morpheme segmentation for 16 languages. Lastly, this new release makes a push towards inclusion of derivational morphology in UniMorph by enriching the data and annotation schema with instances representing derivational processes from MorphyNet.
We release an internationalized annotation and human evaluation bundle, called Textinator, along with documentation and video tutorials. Textinator allows annotating data for a wide variety of NLP tasks, and its user interface is offered in multiple languages, lowering the entry threshold for domain experts. The latter is, in fact, quite a rare feature among the annotation tools, that allows controlling for possible unintended biases introduced due to hiring only English-speaking annotators. We illustrate the rarity of this feature by presenting a thorough systematic comparison of Textinator to previously published annotation tools along 9 different axes (with internationalization being one of them). To encourage researchers to design their human evaluation before starting to annotate data, Textinator offers an easy-to-use tool for human evaluations allowing importing surveys with potentially hundreds of evaluation items in one click. We finish by presenting several use cases of annotation and evaluation projects conducted using pre-release versions of Textinator. The presented use cases do not represent Textinator’s full annotation or evaluation capabilities, and interested readers are referred to the online documentation for more information.
Over the past decades, the number of episodes of cyber aggression occurring online has grown substantially, especially among teens. Most solutions investigated by the NLP community to curb such online abusive behaviors consist of supervised approaches relying on annotated data extracted from social media. However, recent studies have highlighted that private instant messaging platforms are major mediums of cyber aggression among teens. As such interactions remain invisible due to the app privacy policies, very few datasets collecting aggressive conversations are available for the computational analysis of language. In order to overcome this limitation, in this paper we present the CyberAgressionAdo-V1 dataset, containing aggressive multiparty chats in French collected through a role-playing game in high-schools, and annotated at different layers. We describe the data collection and annotation phases, carried out in the context of a EU and a national research projects, and provide insightful analysis on the different types of aggression and verbal abuse depending on the targeted victims (individuals or communities) emerging from the collected data.
There has been a lot of research in identifying hate posts from social media because of their detrimental effects on both individuals and society. The majority of this research has concentrated on English, although one notices the emergence of multilingual detection tools such as multilingual-BERT (mBERT). However, there is a lack of hate speech datasets compared to English, and a multilingual pre-trained model often contains fewer tokens for other languages. This paper attempts to contribute to hate speech identification in Finnish by constructing a new hate speech dataset that is collected from a popular forum (Suomi24). Furthermore, we have experimented with FinBERT pre-trained model performance for Finnish hate speech detection compared to state-of-the-art mBERT and other practices. In addition, we tested the performance of FinBERT compared to fastText as embedding, which employed with Convolution Neural Network (CNN). Our results showed that FinBERT yields a 91.7% accuracy and 90.8% F1 score value, which outperforms all state-of-art models, including multilingual-BERT and CNN.
Automatic post-editing (APE) refers to a research field that aims to automatically correct errors included in the translation sentences derived by the machine translation system. This study has several limitations, considering the data acquisition, because there is no official dataset for most language pairs. Moreover, the amount of data is restricted even for language pairs in which official data has been released, such as WMT. To solve this problem and promote universal APE research regardless of APE data existence, this study proposes a method for automatically generating APE data based on a noising scheme from a parallel corpus. Particularly, we propose a human mimicking errors-based noising scheme that considers a practical correction process at the human level. We propose a precise inspection to attain high performance, and we derived the optimal noising schemes that show substantial effectiveness. Through these, we also demonstrate that depending on the type of noise, the noising scheme-based APE data generation may lead to inferior performance. In addition, we propose a dynamic noise injection strategy that enables the acquisition of a robust error correction capability and demonstrated its effectiveness by comparative analysis. This study enables obtaining a high performance APE model without human-generated data and can promote universal APE research for all language pairs targeting English.
Cross-lingual transfer learning without labeled target language data or parallel text has been surprisingly effective in zero-shot cross-lingual classification, question answering, unsupervised machine translation, etc. However, some recent publications have claimed that domain mismatch prevents cross-lingual transfer, and their results show that unsupervised bilingual lexicon induction (UBLI) and unsupervised neural machine translation (UNMT) do not work well when the underlying monolingual corpora come from different domains (e.g., French text from Wikipedia but English text from UN proceedings). In this work, we show how a simple initialization regimen can overcome much of the effect of domain mismatch in cross-lingual transfer. We pre-train word and contextual embeddings on the concatenated domain-mismatched corpora, and use these as initializations for three tasks: MUSE UBLI, UN Parallel UNMT, and the SemEval 2017 cross-lingual word similarity task. In all cases, our results challenge the conclusions of prior work by showing that proper initialization can recover a large portion of the losses incurred by domain mismatch.
Clinical phenotyping enables the automatic extraction of clinical conditions from patient records, which can be beneficial to doctors and clinics worldwide. However, current state-of-the-art models are mostly applicable to clinical notes written in English. We therefore investigate cross-lingual knowledge transfer strategies to execute this task for clinics that do not use the English language and have a small amount of in-domain data available. Our results reveal two strategies that outperform the state-of-the-art: Translation-based methods in combination with domain-specific encoders and cross-lingual encoders plus adapters. We find that these strategies perform especially well for classifying rare phenotypes and we advise on which method to prefer in which situation. Our results show that using multilingual data overall improves clinical phenotyping models and can compensate for data sparseness.
Translation of the noisy, informal language found in social media has been an understudied problem, with a principal factor being the limited availability of translation corpora in many languages. To address this need we have developed a new corpus containing over 200,000 translations of microblog posts that supports translation of thirteen languages into English. The languages are: Arabic, Chinese, Farsi, French, German, Hindi, Korean, Pashto, Portuguese, Russian, Spanish, Tagalog, and Urdu. We are releasing these data as the Multilingual Microblog Translation Corpus to support futher research in translation of informal language. We establish baselines using this new resource, and we further demonstrate the utility of the corpus by conducting experiments with fine-tuning to improve translation quality from a high performing neural machine translation (NMT) system. Fine-tuning provided substantial gains, ranging from +3.4 to +11.1 BLEU. On average, a relative gain of 21% was observed, demonstrating the utility of the corpus.
Humans constantly deal with multimodal information, that is, data from different modalities, such as texts and images. In order for machines to process information similarly to humans, they must be able to process multimodal data and understand the joint relationship between these modalities. This paper describes the work performed on the VTLM (Visual Translation Language Modelling) framework from (Caglayan et al., 2021) to test its generalization ability for other language pairs and corpora. We use the multimodal and multilingual corpus How2 (Sanabria et al., 2018) in three parallel streams with aligned English-Portuguese-Visual information to investigate the effectiveness of the model for this new language pair and in more complex scenarios, where the sentence associated with each image is not a simple description of it. Our experiments on the Portuguese-English multimodal translation task using the How2 dataset demonstrate the efficacy of cross-lingual visual pretraining. We achieved a BLEU score of 51.8 and a METEOR score of 78.0 on the test set, outperforming the MMT baseline by about 14 BLEU and 14 METEOR. The good BLEU and METEOR values obtained for this new language pair, regarding the original English-German VTLM, establish the suitability of the model to other languages.
Recently, we have seen an increasing interest in the area of speech-to-text translation. This has led to astonishing improvements in this area. In contrast, the activities in the area of speech-to-speech translation is still limited, although it is essential to overcome the language barrier. We believe that one of the limiting factors is the availability of appropriate training data. We address this issue by creating LibriS2S, to our knowledge the first publicly available speech-to-speech training corpus between German and English. For this corpus, we used independently created audio for German and English leading to an unbiased pronunciation of the text in both languages. This allows the creation of a new text-to-speech and speech-to-speech translation model that directly learns to generate the speech signal based on the pronunciation of the source language. Using this created corpus, we propose Text-to-Speech models based on the example of the recently proposed FastSpeech 2 model that integrates source language information. We do this by adapting the model to take information such as the pitch, energy or transcript from the source speech as additional input.
This paper presents a fine-grained test suite for the language pair German–English. The test suite is based on a number of linguistically motivated categories and phenomena and the semi-automatic evaluation is carried out with regular expressions. We describe the creation and implementation of the test suite in detail, providing a full list of all categories and phenomena. Furthermore, we present various exemplary applications of our test suite that have been implemented in the past years, like contributions to the Conference of Machine Translation, the usage of the test suite and MT outputs for quality estimation, and the expansion of the test suite to the language pair Portuguese–English. We describe how we tracked the development of the performance of various systems MT systems over the years with the help of the test suite and which categories and phenomena are prone to resulting in MT errors. For the first time, we also make a large part of our test suite publicly available to the research community.
Recent studies in cross-lingual learning using multilingual models have cast doubt on the previous hypothesis that shared vocabulary and joint pre-training are the keys to cross-lingual generalization. We introduce a method for transferring monolingual models to other languages through continuous pre-training and study the effects of such transfer from four different languages to English. Our experimental results on GLUE show that the transferred models outperform an English model trained from scratch, independently of the source language. After probing the model representations, we find that model knowledge from the source language enhances the learning of syntactic and semantic knowledge in English.
We present a dataset containing source code solutions to algorithmic programming exercises solved by hundreds of Bachelor-level students at the University of Hamburg. These solutions were collected during the winter semesters 2019/2020, 2020/2021 and 2021/2022. The dataset contains a set of solutions to a total of 21 tasks written in Java as well as Python and a total of over 1500 individual solutions. All solutions were submitted through Moodle and the Coderunner plugin and passed a number of test cases (including randomized tests), such that they can be considered as working correctly. All students whose solutions are included in the dataset gave their consent into publishing their solutions. The solutions are pseudonymized with a random solution ID. Included in this paper is a short analysis of the dataset containing statistical data and highlighting a few anomalies (e.g. the number of solutions per task decreases for the last few tasks due to grading rules). We plan to extend the dataset with tasks and solutions from upcoming courses.
In this work, we conduct a quantitative linguistic analysis of the language usage patterns of multilingual peer supporters in two health-focused WhatsApp groups in Kenya comprising of youth living with HIV. Even though the language of communication for the group was predominantly English, we observe frequent use of Kiswahili, Sheng and code-mixing among the three languages. We present an analysis of language choice and its accommodation, different functions of code-mixing, and relationship between sentiment and code-mixing. To explore the effectiveness of off-the-shelf Language Technologies (LT) in such situations, we attempt to build a sentiment analyzer for this dataset. Our experiments demonstrate the challenges of developing LT and therefore effective interventions for such forums and languages. We provide recommendations for language resources that should be built to address these challenges.
Frame shift is a cross-linguistic phenomenon in translation which results in corresponding pairs of linguistic material evoking different frames. The ability to predict frame shifts would enable (semi-)automatic creation of multilingual frame annotations and thus speeding up FrameNet creation through annotation projection. Here, we first characterize how frame shifts result from other linguistic divergences such as translational divergences and construal differences. Our analysis also shows that many pairs of frames in frame shifts are multi-hop away from each other in Berkeley FrameNet’s net-like configuration. Then, we propose the Frame Shift Prediction task and demonstrate that our graph attention networks, combined with auxiliary training, can learn cross-linguistic frame-to-frame correspondence and predict frame shifts.
Cued Speech is a communication system developed for deaf people to complement speechreading at the phonetic level with hands. This visual communication mode uses handshapes in different placements near the face in combination with the mouth movements of speech to make the phonemes of spoken language look different from each other. This paper describes CLeLfPC - Corpus de Lecture en Langue française Parlée Complétée, a corpus of French Cued Speech. It consists in about 4 hours of audio and HD video recordings of 23 participants. The recordings are 160 different isolated ‘CV’ syllables repeated 5 times, 320 words or phrases repeated 2-3 times and about 350 sentences repeated 2-3 times. The corpus is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. It can be used for any further research or teaching purpose. The corpus includes orthographic transliteration and other phonetic annotations on 5 of the recorded topics, i.e. syllables, words, isolated sentences and a text. The early results are encouraging: it seems that 1/ the hand position has a high influence on the key audio duration; and 2/ the hand shape has not.
Samrómur Children is an Icelandic speech corpus intended for the field of automatic speech recognition. It contains 131 hours of read speech from Icelandic children aged between 4 to 17 years. The test portion was meticulously selected to cover a wide range of ages as possible; we aimed to have exactly the same amount of data per age range. The speech was collected with the crowd-sourcing platform Samrómur.is, which is inspired on the “Mozilla’s Common Voice Project”. The corpus was developed within the framework of the “Language Technology Programme for Icelandic 2019 − 2023”; the goal of the project is to make Icelandic available in language-technology applications. Samrómur Children is the first corpus in Icelandic with children’s voices for public use under a Creative Commons license. Additionally, we present baseline experiments and results using Kaldi.
The Norwegian Parliamentary Speech Corpus (NPSC) is a speech dataset with recordings of meetings from Stortinget, the Norwegian parliament. It is the first, publicly available dataset containing unscripted, Norwegian speech designed for training of automatic speech recognition (ASR) systems. The recordings are manually transcribed and annotated with language codes and speakers, and there are detailed metadata about the speakers. The transcriptions exist in both normalized and non-normalized form, and non-standardized words are explicitly marked and annotated with standardized equivalents. To test the usefulness of this dataset, we have compared an ASR system trained on the NPSC with a baseline system trained on only manuscript-read speech. These systems were tested on an independent dataset containing spontaneous, dialectal speech. The NPSC-trained system performed significantly better, with a 22.9% relative improvement in word error rate (WER). Moreover, training on the NPSC is shown to have a “democratizing” effects in terms of dialects, as improvements are generally larger for dialects with higher WER from the baseline system.
We developed a bilingual Frisian/Dutch speech recognizer for council meetings in Fryslân (the Netherlands). During these meetings both Frisian and Dutch are spoken, and code switching between both languages shows up frequently. The new speech recognizer is based on an existing speech recognizer for Frisian and Dutch named FAME!, which was trained and tested on historical radio broadcasts. Adapting a speech recognizer for the council meeting domain is challenging because of acoustic background noise, speaker overlap and the jargon typically used in council meetings. To train the new recognizer, we used the radio broadcast materials utilized for the development of the FAME! recognizer and added newly created manually transcribed audio recordings of council meetings from eleven Frisian municipalities, the Frisian provincial council and the Frisian water board. The council meeting recordings consist of 49 hours of speech, with 26 hours of Frisian speech and 23 hours of Dutch speech. Furthermore, from the same sources, we obtained texts in the domain of council meetings containing 11 million words; 1.1 million Frisian words and 9.9 million Dutch words. We describe the methods used to train the new recognizer, report the observed word error rates, and perform an error analysis on remaining errors.
There is a need for a simple method of detecting early signs of dementia which is not burdensome to patients, since early diagnosis and treatment can often slow the advance of the disease. Several studies have explored using only the acoustic and linguistic information of conversational speech as diagnostic material, with some success. To accelerate this research, we recorded natural conversations between 128 elderly people living in four different regions of Japan and interviewers, who also administered the Hasegawa’s Dementia Scale-Revised (HDS-R), a cognitive impairment test. Using our elderly speech corpus and dementia test results, we propose an SVM-based screening method which can detect dementia using the acoustic features of conversational speech even when regional dialects are present. We accomplish this by omitting some acoustic features, to limit the negative effect of differences between dialects. When using our proposed method, a dementia detection accuracy rate of about 91% was achieved for speakers from two regions. When speech from four regions was used in a second experiment, the discrimination rate fell to 76.6%, but this may have been due to using only sentence-level acoustic features in the second experiment, instead of sentence and phoneme-level features as in the previous experiment. This is an on-going research project, and additional investigation is needed to understand differences in the acoustic characteristics of phoneme units in the conversational speech collected from these four regions, to determine whether the removal of formants and other features can improve the dementia detection rate.
Spoken medical dialogue systems are increasingly attracting interest to enhance access to healthcare services and improve quality and traceability of patient care. In this paper, we focus on medical drug prescriptions acquired on smartphones through spoken dialogue. Such systems would facilitate the traceability of care and would free the clinicians’ time. However, there is a lack of speech corpora to develop such systems since most of the related corpora are in text form and in English. To facilitate the research and development of spoken medical dialogue systems, we present, to the best of our knowledge, the first spoken medical drug prescriptions corpus, named PxNLU. It contains 4 hours of transcribed and annotated dialogues of drug prescriptions in French acquired through an experiment with 55 participants experts and non-experts in prescriptions. We also present some experiments that demonstrate the interest of this corpus for the evaluation and development of medical dialogue systems.
The current largest open-source generic automatic speech recognition (ASR) system for Dutch, Kaldi_NL, does not include a domain-specific healthcare jargon in the lexicon. Commercial alternatives (e.g., Google ASR system) are also not suitable for this purpose, not only because of the lexicon issue, but they do not safeguard privacy of sensitive data sufficiently and reliably. These reasons motivate that just a small amount of medical staff employs speech technology in the Netherlands. This paper proposes an innovative ASR training method developed within the Homo Medicinalis (HoMed) project. On the semantic level it specifically targets automatic transcription of doctor-patient consultation recordings with a focus on the use of medicines. In the first stage of HoMed, the Kaldi_NL language model (LM) is fine-tuned with lists of Dutch medical terms and transcriptions of Dutch online healthcare news bulletins. Despite the acoustic challenges and linguistic complexity of the domain, we reduced the word error rate (WER) by 5.2%. The proposed method could be employed for ASR domain adaptation to other domains with sensitive and special category data. These promising results allow us to apply this methodology on highly sensitive audiovisual recordings of patient consultations at the Netherlands Institute for Health Services Research (Nivel).
Machine learning methodologies can be adopted in cultural applications and propose new ways to distribute or even present the cultural content to the public. For instance, speech analytics can be adopted to automatically generate subtitles in theatrical plays, in order to (among other purposes) help people with hearing loss. Apart from a typical speech-to-text transcription with Automatic Speech Recognition (ASR), Speech Emotion Recognition (SER) can be used to automatically predict the underlying emotional content of speech dialogues in theatrical plays, and thus to provide a deeper understanding how the actors utter their lines. However, real-world datasets from theatrical plays are not available in the literature. In this work we present GreThE, the Greek Theatrical Emotion dataset, a new publicly available data collection for speech emotion recognition in Greek theatrical plays. The dataset contains utterances from various actors and plays, along with respective valence and arousal annotations. Towards this end, multiple annotators have been asked to provide their input for each speech recording and inter-annotator agreement is taken into account in the final ground truth generation. In addition, we discuss the results of some indicative experiments that have been conducted with machine and deep learning frameworks, using the dataset, along with some widely used databases in the field of speech emotion recognition.
Synthetic voices are increasingly used in applications that require a conversational speaking style, raising the question as to which type of training data yields the most suitable speaking style for such applications. This study compares voices trained on three corpora of equal size recorded by the same speaker: an audiobook character speech (dialogue) corpus, an audiobook narrator speech corpus, and a neutral-style sentence-based corpus. The voices were trained with three text-to-speech synthesisers: two hidden Markov model-based synthesisers and a neural synthesiser. An evaluation study tested the suitability of their speaking style for use in customer service voice chatbots. Independently of the synthesiser used, the voices trained on the character speech corpus received the lowest, and those trained on the neutral-style corpus the highest scores. However, the evaluation results may have been confounded by the greater acoustic variability, less balanced sentence length distribution, and poorer phonemic coverage of the character speech corpus, especially compared to the neutral-style corpus. Therefore, the next step will be the creation of a more uniform, balanced, and representative audiobook dialogue corpus, and the evaluation of its suitability for further conversational-style applications besides customer service chatbots.
Speech characteristics vary from speaker to speaker. While some variation phenomena are due to the overall communication setting, others are due to diastratic factors such as gender, provenance, age, and social background. The analysis of these factors, although relevant for both linguistic and speech technology communities, is hampered by the need to annotate existing corpora or to recruit, categorise, and record volunteers as a function of targeted profiles. This paper presents a methodology that uses a knowledge base to provide speaker-specific information. This can facilitate the enrichment of existing corpora with new annotations extracted from the knowledge base. The method also helps the large scale analysis by automatically extracting instances of speech variation to correlate with diastratic features. We apply our method to an over 120-hour corpus of broadcast speech in French and investigate variation patterns linked to reduction phenomena and/or specific to connected speech such as disfluencies. We find significant differences in speech rate, the use of filler words, and the rate of non-canonical realisations of frequent segments as a function of different professional categories and age groups.
Identifying phone inventories is a crucial component in language documentation and the preservation of endangered languages. However, even the largest collection of phone inventory only covers about 2000 languages, which is only 1/4 of the total number of languages in the world. A majority of the remaining languages are endangered. In this work, we attempt to solve this problem by estimating the phone inventory for any language listed in Glottolog, which contains phylogenetic information regarding 8000 languages. In particular, we propose one probabilistic model and one non-probabilistic model, both using phylogenetic trees (“language family trees”) to measure the distance between languages. We show that our best model outperforms baseline models by 6.5 F1. Furthermore, we demonstrate that, with the proposed inventories, the phone recognition model can be customized for every language in the set, which improved the PER (phone error rate) in phone recognition by 25%.
This paper presents a collection of parallel corpora generated by exploiting the COVID-19 related dataset of metadata created with the Europe Media Monitor (EMM) / Medical Information System (MediSys) processing chain of news articles. We describe how we constructed comparable monolingual corpora of news articles related to the current pandemic and used them to mine about 11.2 million segment alignments in 26 EN-X language pairs, covering most official EU languages plus Albanian, Arabic, Icelandic, Macedonian, and Norwegian. Subsets of this collection have been used in shared tasks (e.g. Multilingual Semantic Search, Machine Translation) aimed at accelerating the creation of resources and tools needed to facilitate access to information in the COVID-19 emergency situation.
We introduce ChemDisGene, a new dataset for training and evaluating multi-class multi-label biomedical relation extraction models. Our dataset contains 80k biomedical research abstracts labeled with mentions of chemicals, diseases, and genes, portions of which human experts labeled with 18 types of biomedical relationships between these entities (intended for evaluation), and the remainder of which (intended for training) has been distantly labeled via the CTD database with approximately 78% accuracy. In comparison to similar preexisting datasets, ours is both substantially larger and cleaner; it also includes annotations linking mentions to their entities. We also provide three baseline deep neural network relation extraction models trained and evaluated on our new dataset.
This paper develops the first question answering dataset (DrugEHRQA) containing question-answer pairs from both structured tables and unstructured notes from a publicly available Electronic Health Record (EHR). EHRs contain patient records, stored in structured tables and unstructured clinical notes. The information in structured and unstructured EHRs is not strictly disjoint: information may be duplicated, contradictory, or provide additional context between these sources. Our dataset has medication-related queries, containing over 70,000 question-answer pairs. To provide a baseline model and help analyze the dataset, we have used a simple model (MultimodalEHRQA) which uses the predictions of a modality selection network to choose between EHR tables and clinical notes to answer the questions. This is used to direct the questions to the table-based or text-based state-of-the-art QA model. In order to address the problem arising from complex, nested queries, this is the first time Relation-Aware Schema Encoding and Linking for Text-to-SQL Parsers (RAT-SQL) has been used to test the structure of query templates in EHR data. Our goal is to provide a benchmark dataset for multi-modal QA systems, and to open up new avenues of research in improving question answering over EHR structured data by using context from unstructured clinical data.
Neural Network (NN) architectures are used more and more to model large amounts of data, such as text data available online. Transformer-based NN architectures have shown to be very useful for language modelling. Although many researchers study how such Language Models (LMs) work, not much attention has been paid to the privacy risks of training LMs on large amounts of data and publishing them online. This paper presents a new method for anonymizing a language model by presenting the way in which MedRoBERTa.nl, a Dutch language model for hospital notes, was anonymized. The two-step method involves i) automatic anonymization of the training data and ii) semi-automatic anonymization of the LM’s vocabulary. Adopting the fill-mask task where the model predicts what tokens are most probable in a certain context, it was tested how often the model will predict a name in a context where a name should be. It was shown that it predicts a name-like token 0.2% of the time. Any name-like token that was predicted was never the name originally present in the training data. By explaining how a LM trained on highly private real-world medical data can be published, we hope that more language resources will be published openly and responsibly so the scientific community can profit from them.
The successes of contextual word embeddings learned by training large-scale language models, while remarkable, have mostly occurred for languages where significant amounts of raw texts are available and where annotated data in downstream tasks have a relatively regular spelling. Conversely, it is not yet completely clear if these models are also well suited for lesser-resourced and more irregular languages. We study the case of Old French, which is in the interesting position of having relatively limited amount of available raw text, but enough annotated resources to assess the relevance of contextual word embedding models for downstream NLP tasks. In particular, we use POS-tagging and dependency parsing to evaluate the quality of such models in a large array of configurations, including models trained from scratch from small amounts of raw text and models pre-trained on other languages but fine-tuned on Medieval French data.
The prevailing practice in the academia is to evaluate the model performance on in-domain evaluation data typically set aside from the training corpus. However, in many real world applications the data on which the model is applied may very substantially differ from the characteristics of the training data. In this paper, we focus on Finnish out-of-domain parsing by introducing a novel UD Finnish-OOD out-of-domain treebank including five very distinct data sources (web documents, clinical, online discussions, tweets, and poetry), and a total of 19,382 syntactic words in 2,122 sentences released under the Universal Dependencies framework. Together with the new treebank, we present extensive out-of-domain parsing evaluation utilizing the available section-level information from three different Finnish UD treebanks (TDT, PUD, OOD). Compared to the previously existing treebanks, the new Finnish-OOD is shown include sections more challenging for the general parser, creating an interesting evaluation setting and yielding valuable information for those applying the parser outside of its training domain.
In this paper we present the final result of a project focused on Tunisian Arabic encoded in Arabizi, the Latin-based writing system for digital conversations. The project led to the realization of two integrated and independent tools: a linguistic corpus and a neural network architecture created to annotate the former with various levels of linguistic information (code-switching classification, transliteration, tokenization, POS-tagging, lemmatization). We discuss the choices made in terms of computational and linguistic methodology and the strategies adopted to improve our results. We report on the experiments performed in order to outline our research path. Finally, we explain the reasons why we believe in the potential of these tools for both computational and linguistic researches.
Our work aims at developing a multilingual data resource for morphological segmentation. We present a survey of 17 existing data resources relevant for segmentation in 32 languages, and analyze diversity of how individual linguistic phenomena are captured across them. Inspired by the success of Universal Dependencies, we propose a harmonized scheme for segmentation representation, and convert the data from the studied resources into this common scheme. Harmonized versions of resources available under free licenses are published as a collection called UniSegments 1.0.
We present the TeDDi sample, a diversity sample of text data for language comparison and multilingual Natural Language Processing. The TeDDi sample currently features 89 languages based on the typological diversity sample in the World Atlas of Language Structures. It consists of more than 20k texts and is accompanied by open-source corpus processing tools. The aim of TeDDi is to facilitate text-based quantitative analysis of linguistic diversity. We describe in detail the TeDDi sample, how it was created, data availability, and its added value through for NLP and linguistic research.
Word embeddings (Mikolov et al., 2013; Pennington et al., 2014) have been used to bolster the performance of natural language processing systems in a wide variety of tasks, including information retrieval (Roy et al., 2018) and machine translation (Qi et al., 2018). However, approaches to learning word embeddings typically require large corpora of running text to learn high quality representations. For many languages, such resources are unavailable. This is the case for Wolastoqey, also known as Passamaquoddy-Maliseet, an endangered low-resource Indigenous language. As there exist no large corpora of running text for Wolastoqey, in this paper, we leverage a bilingual dictionary to learn Wolastoqey word embeddings by encoding their corresponding English definitions into vector representations using pretrained English word and sequence representation models. Specifically, we consider representations based on pretrained word2vec (Mikolov et al., 2013), RoBERTa (Liu et al., 2019) and sentence-BERT (Reimers and Gurevych, 2019) models. We evaluate these embeddings in word prediction tasks focused on part-of-speech, animacy, and transitivity; semantic clustering; and reverse dictionary search. In all evaluations we demonstrate that approaches using these embeddings outperform task-specific baselines, without requiring any language-specific training or fine-tuning.
Machine learning (ML) approaches have dominated NLP during the last two decades. From machine translation and speech technology, ML tools are now also in use for spellchecking and grammar checking, with a blurry distinction between the two. We unmask the myth of effortless big data by illuminating the efforts and time that lay behind building a multi-purpose corpus with regard to collecting, mark-up and building from scratch. We also discuss what kind of language technology minority languages actually need, and to what extent the dominating paradigm has been able to deliver these tools. In this context we present our alternative to corpus-based language technology, which is knowledge-based language technology, and we show how this approach can provide language technology solutions for languages being outside the reach of machine learning procedures. We present a stable and mature infrastructure (GiellaLT) containing more than hundred languages and building a number of language technology tools that are useful for language communities.
We present an analysis pipeline and best practice guidelines for building and curating corpora of everyday conversation in diverse languages. Surveying language documentation corpora and other resources that cover 67 languages and varieties from 28 phyla, we describe the compilation and curation process, specify minimal properties of a unified format for interactional data, and develop methods for quality control that take into account turn-taking and timing. Two case studies show the broad utility of conversational data for (i) charting human interactional infrastructure and (ii) tracing challenges and opportunities for current ASR solutions. Linguistically diverse conversational corpora can provide new insights for the language sciences and stronger empirical foundations for language technology.
In this paper, we describe the creation and annotation of EPIC UdS, a multilingual corpus of simultaneous interpreting for English, German and Spanish. We give an overview of the comparable and parallel, aligned corpus variants and explore various applications of the corpus. What makes EPIC UdS relevant is that it is one of the rare interpreting corpora that includes transcripts suitable for research on more than one language pair and on interpreting with regard to German. It not only contains transcribed speeches, but also rich metadata and fine-grained linguistic annotations tailored for diverse applications across a broad range of linguistic subfields.
We present the development of a benchmark suite consisting of an annotation schema, training corpus and baseline model for Entity Recognition (ER) in job descriptions, published under a Creative Commons license. This was created to address the distinct lack of resources available to the community for the extraction of salient entities, such as skills, from job descriptions. The dataset contains 18.6k entities comprising five types (Skill, Qualification, Experience, Occupation, and Domain). We include a benchmark CRF-based ER model which achieves an F1 score of 0.59. Through the establishment of a standard definition of entities and training/testing corpus, the suite is designed as a foundation for future work on tasks such as the development of job recommender systems.
CAMIO (Corpus of Annotated Multilingual Images for OCR) is a new corpus created by Linguistic Data Consortium to serve as a resource to support the development and evaluation of optical character recognition (OCR) and related technologies for 35 languages across 24 unique scripts. The corpus comprises nearly 70,000 images of machine printed text, covering a wide variety of topics and styles, document domains, attributes and scanning/capture artifacts. Most images have been exhaustively annotated for text localization, resulting in over 2.3M line-level bounding boxes. For 13 of the 35 languages, 1250 images/language have been further annotated with orthographic transcriptions of each line plus specification of reading order, yielding over 2.4M tokens of transcribed text. The resulting annotations are represented in a comprehensive XML output format defined for this corpus. The paper discusses corpus design and implementation, challenges encountered, baseline performance results obtained on the corpus for text localization and OCR decoding, and plans for corpus publication.
In this paper, we present the FABRA: readability toolkit based on the aggregation of a large number of readability predictor variables. The toolkit is implemented as a service-oriented architecture, which obviates the need for installation, and simplifies its integration into other projects. We also perform a set of experiments to show which features are most predictive on two different corpora, and how the use of aggregators improves performance over standard feature-based readability prediction. Our experiments show that, for the explored corpora, the most important predictors for native texts are measures of lexical diversity, dependency counts and text coherence, while the most important predictors for foreign texts are syntactic variables illustrating language development, as well as features linked to lexical sophistication. FABRA: have the potential to support new research on readability assessment for French.
Speech interfaces for argumentative dialogue systems (ADS) are rather scarce. The complex task they pursue hinders the application of common natural language understanding (NLU) approaches in this domain. To address this issue we include an adaption of a recently introduced NLU framework tailored to argumentative tasks into a complete ADS. We evaluate the likeability and motivation of users to interact with the new system in a user study. Therefore, we compare it to a solid baseline utilizing a drop-down menu. The results indicate that the integration of a flexible NLU framework enables a far more natural and satisfying interaction with human users in real-time. Even though the drop-down menu convinces regarding its robustness, the willingness to use the new system is significantly higher. Hence, the featured NLU framework provides a sound basis to build an intuitive interface which can be extended to adapt its behavior to the individual user.
We propose a deep learning-based foreign language learning platform, named FreeTalky, for people who experience anxiety dealing with foreign languages, by employing a humanoid robot NAO and various deep learning models. A persona-based dialogue system that is embedded in NAO provides an interesting and consistent multi-turn dialogue for users. Also, an grammar error correction system promotes improvement in grammar skills of the users. Thus, our system enables personalized learning based on persona dialogue and facilitates grammar learning of a user using grammar error feedback. Furthermore, we verified whether FreeTalky provides practical help in alleviating xenoglossophobia by replacing the real human in the conversation with a NAO robot, through human evaluation.
Often both an utterance and its context must be read to understand its intent in a dialog. Herein we propose a task, Self- Contained Utterance Description (SCUD), to describe the intent of an utterance in a dialog with multiple simple natural sentences without the context. If a task can be performed concurrently with high accuracy as the conversation continues such as in an accommodation search dialog, the operator can easily suggest candidates to the customer by inputting SCUDs of the customer’s utterances to the accommodation search system. SCUDs can also describe the transition of customer requests from the dialog log. We construct a Japanese corpus to train and evaluate automatic SCUD generation. The corpus consists of 210 dialogs containing 10,814 sentences. We conduct an experiment to verify that SCUDs can be automatically generated. Additionally, we investigate the influence of the amount of training data on the automatic generation performance using 8,200 additional examples.
Dialog system developers need high-quality data to train, fine-tune and assess their systems. They often use crowdsourcing for this since it provides large quantities of data from many workers. However, the data may not be of sufficiently good quality. This can be due to the way that the requester presents a task and how they interact with the workers. This paper introduces DialCrowd 2.0 to help requesters obtain higher quality data by, for example, presenting tasks more clearly and facilitating effective communication with workers. DialCrowd 2.0 guides developers in creating improved Human Intelligence Tasks (HITs) and is directly applicable to the workflows used currently by developers and researchers.
Several dialogue corpora are currently available for research purposes, but they still fall short for the growing interest in the development of dialogue systems with their own specific requirements. In order to help those requiring such a corpus, this paper surveys a range of available options, in terms of aspects like speakers, size, languages, collection, annotations, and domains. Some trends are identified and possible approaches for the creation of new corpora are also discussed.
In human-human conversations, Context Tracking deals with identifying important entities and keeping track of their properties and relationships. This is a challenging problem that encompasses several subtasks such as slot tagging, coreference resolution, resolving plural mentions and entity linking. We approach this problem as an end-to-end modeling task where the conversational context is represented by an entity repository containing the entity references mentioned so far, their properties and the relationships between them. The repository is updated turn-by-turn, thus making training and inference computationally efficient even for long conversations. This paper lays the groundwork for an investigation of this framework in two ways. First, we release Contrack, a large scale human-human conversation corpus for context tracking with people and location annotations. It contains over 7000 conversations with an average of 11.8 turns, 5.8 entities and 15.2 references per conversation. Second, we open-source a neural network architecture for context tracking. Finally we compare this network to state-of-the-art approaches for the subtasks it subsumes and report results on the involved tradeoffs.
Every model is only as strong as the data that it is trained on. In this paper, we present a new dataset, obtained by merging four publicly available annotated corpora for task-oriented dialogues in several domains (MultiWOZ 2.2, CamRest676, DSTC2 and Schema-Guided Dialogue Dataset). This way, we assess the feasibility of providing a unified ontology and annotation schema covering several domains with a relatively limited effort. We analyze the characteristics of the resulting dataset along three main dimensions: language, information content and performance. We focus on aspects likely to be pertinent for improving dialogue success, e.g. dialogue consistency. Furthermore, to assess the usability of this new corpus, we thoroughly evaluate dialogue generation performance under various conditions with the help of two prominent recent end-to-end dialogue models: MarCo and GPT-2. These models were selected as popular open implementations representative of the two main dimensions of dialogue modelling. While we did not observe a significant gain for dialogue state tracking performance, we show that using more training data from different sources can improve language modelling capabilities and positively impact dialogue flow (consistency). In addition, we provide the community with one of the largest open dataset for machine learning experiments.
In dialogue analysis, characterising named entities in the domain of interest is relevant in order to understand how people are making use of them for argumentation purposes. The movie recommendation domain is a frequently considered case study for many applications and by linguistic studies and, since many different resources have been collected throughout the years to describe it, a single database combining all these data sources is a valuable asset for cross-disciplinary investigations. We propose an integrated graph-based structure of multiple resources, enriched with the results of the application of graph analytics approaches to provide an encompassing view of the domain and of the way people talk about it during the recommendation task. While we cannot distribute the final resource because of licensing issues, we share the code to assemble and process it once the reference data have been obtained from the original sources.
In this paper we present SHARE, a new lexical resource with 10,125 offensive terms and expressions collected from Spanish speakers. We retrieve this vocabulary using an existing chatbot developed to engage a conversation with users and collect insults via Telegram, named Fiero. This vocabulary has been manually labeled by five annotators obtaining a kappa coefficient agreement of 78.8%. In addition, we leverage the lexicon to release the first corpus in Spanish for offensive span identification research named OffendES_spans. Finally, we show the utility of our resource as an interpretability tool to explain why a comment may be considered offensive.
We present a machine-readable structured data version of Wiktionary. Unlike previous Wiktionary extractions, the new extractor, Wiktextract, fully interprets and expands templates and Lua modules in Wiktionary. This enables it to perform a more complete, robust, and maintainable extraction. The extracted data is multilingual and includes lemmas, inflected forms, translations, etymology, usage examples, pronunciations (including URLs of sound files), lexical and semantic relations, and various morphological, syntactic, semantic, topical, and dialectal annotations. We extract all data from the English Wiktionary. Comparing against previous extractions from language-specific dictionaries, we find that its coverage for non-English languages often matches or exceeds the coverage in the language-specific editions, with the added benefit that all glosses are in English. The data is freely available and regularly updated, enabling anyone to add more data and correct errors by editing Wiktionary. The extracted data is in JSON format and designed to be easy to use by researchers, downstream resources, and application developers.
What makes a text easy to read or not, depends on a variety of factors. One of the most prominent is, however, if the text contains easy, and avoids difficult, words. Deciding if a word is easy or difficult is not a trivial task, since it depends on characteristics of the word in itself as well as the reader, but it can be facilitated by the help of a corpus annotated with word frequencies and reading proficiency levels. In this paper, we present NyLLex, a novel lexical resource derived from books published by Sweden’s largest publisher for easy language texts. NyLLex consists of 6,668 entries, with frequency counts distributed over six reading proficiency levels. We show that NyLLex, with its novel source material aimed at individuals of different reading proficiency levels, can serve as a complement to already existing resources for Swedish.
We present an extension of the SynSemClass Event-type Ontology, originally conceived as a bilingual Czech-English resource. We added German entries to the classes representing the concepts of the ontology. Having a different starting point than the original work (unannotated parallel corpus without links to a valency lexicon and, of course, different existing lexical resources), it was a challenge to adapt the annotation guidelines, the data model and the tools used for the original version. We describe the process and results of working in such a setup. We also show the next steps to adapt the annotation process, data structures and formats and tools necessary to make the addition of a new language in the future more smooth and efficient, and possibly to allow for various teams to work on SynSemClass extensions to many languages concurrently. We also present the latest release which contains the results of adding German, freely available for download as well as for online access.
We present NomVallex, a manually annotated valency lexicon of Czech nouns and adjectives. The lexicon is created in the theoretical framework of the Functional Generative Description and based on corpus data. In total, NomVallex 2.0 is comprised of 1027 lexical units contained in 570 lexemes, covering the following part-of-speech and derivational categories: deverbal and deadjectival nouns, and deverbal, denominal, deadjectival and primary adjectives. Valency properties of a lexical unit are captured in a valency frame which is modeled as a sequence of valency slots, supplemented with a list of morphemic forms. In order to make it possible to study the relationship between valency behavior of base words and their derivatives, lexical units of nouns and adjectives in NomVallex are linked to their respective base words, contained either in NomVallex itself or, in case of verbs, in a valency lexicon of Czech verbs called VALLEX. NomVallex enables a comparison of valency properties of a significant number of Czech nominals with their base words, both manually and in an automatic way; as such, we can address the theoretical question of argument inheritance, concentrating on systemic and non-systemic valency behavior.
Terminology databases are highly useful for the dissemination of specialized knowledge. In this paper we present TZOS, an online terminology database to work on Basque academic terminology collaboratively. We show how this resource integrates the Communicative Theory of Terminology, together with the methodological matters, how it is connected with real corpus GARATERM, which terminology issues arise when terms are collected and future perspectives. The main objectives of this work are to develop basic tools to research academic registers and make the terminology collected by expert users available to the community. Even though TZOS has been designed for an educational context, its flexible structure makes possible to extend it also to the professional area. In this way, we have built IZIBI-TZOS which is a Civil Engineering oriented version of TZOS. These resources are already publicly available, and the ongoing work is towards the interlinking with other lexical resources by applying linking data principles.
In this paper, we introduce a gold standard for animacy detection comprising almost 14,500 German nouns that might be used to denote either animate entities or non-animate entities. We present inter-annotator agreement of our crowd-sourced seed annotations (9,000 nouns) and discuss the results of machine learning models applied to this data.
Emotion classification is often formulated as the task to categorize texts into a predefined set of emotion classes. So far, this task has been the recognition of the emotion of writers and readers, as well as that of entities mentioned in the text. We argue that a classification setup for emotion analysis should be performed in an integrated manner, including the different semantic roles that participate in an emotion episode. Based on appraisal theories in psychology, which treat emotions as reactions to events, we compile an English corpus of written event descriptions. The descriptions depict emotion-eliciting circumstances, and they contain mentions of people who responded emotionally. We annotate all experiencers, including the original author, with the emotions they likely felt. In addition, we link them to the event they found salient (which can be different for different experiencers in a text) by annotating event properties, or appraisals (e.g., the perceived event undesirability, the uncertainty of its outcome). Our analysis reveals patterns in the co-occurrence of people’s emotions in interaction. Hence, this richly-annotated resource provides useful data to study emotions and event evaluations from the perspective of different roles, and it enables the development of experiencer-specific emotion and appraisal classification systems.
In this paper, we discuss work that strives to measure the degree of negativity - the negative polar load - of noun phrases, especially those denoting actors. Since no gold standard data is available for German for this quantification task, we generated a silver standard and used it to fine-tune a BERT-based intensity regressor. We evaluated the quality of the silver standard empirically and found that our lexicon-based quantification metric showed a strong correlation with human annotators.
In this paper, we introduce a new Czech subjectivity dataset of 10k manually annotated subjective and objective sentences from movie reviews and descriptions. Our prime motivation is to provide a reliable dataset that can be used with the existing English dataset as a benchmark to test the ability of pre-trained multilingual models to transfer knowledge between Czech and English and vice versa. Two annotators annotated the dataset reaching 0.83 of the Cohen’s K inter-annotator agreement. To the best of our knowledge, this is the first subjectivity dataset for the Czech language. We also created an additional dataset that consists of 200k automatically labeled sentences. Both datasets are freely available for research purposes. Furthermore, we fine-tune five pre-trained BERT-like models to set a monolingual baseline for the new dataset and we achieve 93.56% of accuracy. We fine-tune models on the existing English dataset for which we obtained results that are on par with the current state-of-the-art results. Finally, we perform zero-shot cross-lingual subjectivity classification between Czech and English to verify the usability of our dataset as the cross-lingual benchmark. We compare and discuss the cross-lingual and monolingual results and the ability of multilingual models to transfer knowledge between languages.
RED (Romanian Emotion Dataset) is a machine learning-based resource developed for the automatic detection of emotions in Romanian texts, containing single-label annotated tweets with one of the following emotions: joy, fear, sadness, anger and neutral. In this work, we propose REDv2, an open-source extension of RED by adding two more emotions, trust and surprise, and by widening the annotation schema so that the resulted novel dataset is multi-label. We show the overall reliability of our dataset by computing inter-annotator agreements per tweet using a formula suitable for our annotation setup and we aggregate all annotators’ opinions into two variants of ground truth, one suitable for multi-label classification and the other suitable for text regression. We propose strong baselines with two transformer models, the Romanian BERT and the multilingual XLM-Roberta model, in both categorical and regression settings.
The traditional evaluation of labeled spans with precision, recall, and F1-score has undesirable effects due to double penalties. Annotations with incorrect label or boundaries count as two errors instead of one, despite being closer to the target annotation than false positives or false negatives. In this paper, new error types are introduced, which more accurately reflect true annotation quality and ensure that every annotation counts only once. An algorithm for error identification in flat and multi-level annotations is presented and complemented with a proposal on how to calculate meaningful precision, recall, and F1-scores based on the more fine-grained error types. The exemplary application to three different annotation tasks (NER, chunking, parsing) shows that the suggested procedure not only prevents double penalties but also allows for a more detailed error analysis, thereby providing more insight into the actual weaknesses of a system.
Multiple works have proposed to probe language models (LMs) for generalization in named entity (NE) typing (NET) and recognition (NER). However, little has been done in this direction for auto-regressive models despite their popularity and potential to express a wide variety of NLP tasks in the same unified format. We propose a new methodology to probe auto-regressive LMs for NET and NER generalization, which draws inspiration from human linguistic behavior, by resorting to meta-learning. We study NEs of various types individually by designing a zero-shot transfer strategy for NET. Then, we probe the model for NER by providing a few examples at inference. We introduce a novel procedure to assess the model’s memorization of NEs and report the memorization’s impact on the results. Our findings show that: 1) GPT2, a common pre-trained auto-regressive LM, without any fine-tuning for NET or NER, performs the tasksfairly well; 2) name irregularity when common for a NE type could be an effective exploitable cue; 3) the model seems to rely more on NE than contextual cues in few-shot NER; 4) NEs with words absent during LM pre-training are very challenging for both NET and NER.
As input representation for each sub-word, the original BERT architecture proposes the sum of the sub-word embedding, position embedding and a segment embedding. Sub-word and position embeddings are well-known and studied, and encode lexical information and word position, respectively. In contrast, segment embeddings are less known and have so far received no attention, despite being ubiquitous in large pre-trained language models. The key idea of segment embeddings is to encode to which of the two sentences (segments) a word belongs to — the intuition is to inform the model about the separation of sentences for the next sentence prediction pre-training task. However, little is known on whether the choice of segment impacts performance. In this work, we try to fill this gap and empirically study the impact of the segment embedding during inference time for a variety of pre-trained embeddings and target tasks. We hypothesize that for single-sentence prediction tasks performance is not affected — neither in mono- nor multilingual setups — while it matters when swapping segment IDs in paired-sentence tasks. To our surprise, this is not the case. Although for classification tasks and monolingual BERT models no large differences are observed, particularly word-level multilingual prediction tasks are heavily impacted. For low-resource syntactic tasks, we observe impacts of segment embedding and multilingual BERT choice. We find that the default setting for the most used multilingual BERT model underperforms heavily, and a simple swap of the segment embeddings yields an average improvement of 2.5 points absolute LAS score for dependency parsing over 9 different treebanks.
This paper addresses the semi-automatic annotation of subjects, also called policy areas, in the Danish Parliament Corpus (2009-2017) v.2. Recently, the corpus has been made available through the CLARIN-DK repository, the Danish node of the European CLARIN infrastructure. The paper also contains an analysis of the subjects in the corpus, and a description of multi-label classification experiments act to verify the consistency of the subject annotation and the utility of the corpus for training classifiers on this type of data. The analysis of the corpus comprises an investigation of how often the parliament members addressed each subject and the relation between subjects and gender of the speaker. The classification experiments show that classifiers can determine the two co-occurring subjects of the speeches from the agenda titles with a performance similar to that of human annotators. Moreover, a multilayer perceptron achieved an F1-score of 0.68 on the same task when trained on bag of words vectors obtained from the speeches’ lemmas. This is an improvement of more than 0.6 with respect to the baseline, a majority classifier that accounts for the frequency of the classes. The result is promising given the high number of subject combinations (186) and the skewness of the data.
A significant challenge in developing translation systems for the world’s ∼7,000 languages is that very few have sufficient data for state-of-the-art techniques. Transfer learning is a promising direction for low-resource neural machine translation (NMT), but introduces many new variables which are often selected through ablation studies, costly trial-and-error, or niche expertise. When pre-training an NMT system for low-resource translation, the pre-training task is often chosen based on data abundance and similarity to the main task. Factors such as dataset sizes and similarity have typically been analysed independently in previous studies, due to the computational cost associated with systematic studies. However, these factors are not independent. We conducted a three-factor experiment to examine how language similarity, pre-training dataset size and main dataset size interacted in their effect on performance in pre-trained transformer-based low-resource NMT. We replicated the common finding that more data was beneficial in bilingual systems, but also found a statistically significant interaction between the three factors, which reduced the effectiveness of large pre-training datasets for some main task dataset sizes (p-value < 0.0018). The surprising trends identified in these interactions indicate that systematic studies of interactions may be a promising long-term direction for guiding research in low-resource neural methods.
TimeML is a scheme for representing temporal information (times, events, & temporal relations) in texts. Although automatic TimeML annotation is challenging, there has been notable progress, with F1s of 0.8–0.9 for events and time detection subtasks, and F1s of 0.5–0.7 for relation extraction. Individually, these subtask results are reasonable, even good, but when combined to generate a full TimeML graph, is overall performance still acceptable? We present a novel suite of eight metrics, combined with a new graph-transformation experimental design, for holistic evaluation of TimeML graphs. We apply these metrics to four automatic TimeML annotation systems (CAEVO, TARSQI, CATENA, and ClearTK). We show that on average 1/3 of the TimeML graphs produced using these systems are inconsistent, and there is on average 1/5 more temporal indeterminacy than the gold-standard. We also show that the automatically generated graphs are on average 109 edits from the gold-standard, which is 1/3 toward complete replacement. Finally, we show that the relationship individual subtask performance and graph quality is non-linear: small errors in TimeML subtasks result in rapid degradation of final graph quality. These results suggest current automatic TimeML annotators are far from optimal and significant further improvement would be useful.
From both human translators (HT) and machine translation (MT) researchers’ point of view, translation quality evaluation (TQE) is an essential task. Translation service providers (TSPs) have to deliver large volumes of translations which meet customer specifications with harsh constraints of required quality level in tight time-frames and costs. MT researchers strive to make their models better, which also requires reliable quality evaluation. While automatic machine translation evaluation (MTE) metrics and quality estimation (QE) tools are widely available and easy to access, existing automated tools are not good enough, and human assessment from professional translators (HAP) are often chosen as the golden standard (CITATION). Human evaluations, however, are often accused of having low reliability and agreement. Is this caused by subjectivity or statistics is at play? How to avoid the entire text to be checked and be more efficient with TQE from cost and efficiency perspectives, and what is the optimal sample size of the translated text, so as to reliably estimate the translation quality of the entire material? This work carries out such a motivated research to correctly estimate the confidence intervals (CITATION) depending on the sample size of translated text, e.g. the amount of words or sentences, that needs to be processed on TQE workflow step for confident and reliable evaluation of overall translation quality. The methodology we applied for this work is from Bernoulli Statistical Distribution Modelling (BSDM) and Monte Carlo Sampling Analysis (MCSA).
Transformer-based models showed near-perfect results on several downstream tasks. However, their performance on classical Arabic texts is largely unexplored. To fill this gap, we evaluate monolingual, bilingual, and multilingual state-of-the-art models to detect relatedness between the Quran (Muslim holy book) and the Hadith (Prophet Muhammed teachings), which are complex classical Arabic texts with underlying meanings that require deep human understanding. To do this, we carefully built a dataset of Quran-verse and Hadith-teaching pairs by consulting sources of reputable religious experts. This study presents the methodology of creating the dataset, which we make available on our repository, and discusses the models’ performance that calls for the imminent need to explore avenues for improving the quality of these models to capture the semantics in such complex, low-resource texts.
Visual Question Answering (VQA) is a challenge problem that can advance AI by integrating several important sub-disciplines including natural language understanding and computer vision. Large VQA datasets that are publicly available for training and evaluation have driven the growth of VQA models that have obtained increasingly larger accuracy scores. However, it is also important to understand how much a model understands the details that are provided in a question. For example, studies in psychology have shown that syntactic complexity places a larger cognitive load on humans. Analogously, we want to understand if models have the perceptual capability to handle modifications to questions. Therefore, we develop a new dataset using Amazon Mechanical Turk where we asked workers to add modifiers to questions based on object properties and spatial relationships. We evaluate this data on LXMERT which is a state-of-the-art model in VQA that focuses more extensively on language processing. Our conclusions indicate that there is a significant negative impact on the performance of the model when the questions are modified to include more detailed information.
The paper presents the outcomes of AI-COVID19, our project aimed at better understanding of misinformation flow about COVID-19 across social media platforms. The specific focus of the study reported in this paper is on collecting data from Telegram groups which are active in promotion of COVID-related misinformation. Our corpus collected so far contains around 28 million words, from almost one million messages. Given that a substantial portion of misinformation flow in social media is spread via multimodal means, such as images and video, we have also developed a mechanism for utilising such channels via producing automatic transcripts for videos and automatic classification for images into such categories as memes, screenshots of posts and other kinds of images. The accuracy of the image classification pipeline is around 87%.
In this study, we thrive on finding out how code-switching and code-mixing (CS/CM) as a linguistic phenomenon could be a sign of tension in Holocaust survivors’ interviews. We first created an interview corpus (a total of 39 interviews) that contains manually annotated CS/CM codes (a total of 802 quotations). We then compared our annotations with the tension places in the corpus. The tensions are identified by a computational tool. We found that most of our annotations were captured in the tension places, and it showed a relatively outstanding performance. The finding implies that CS/CM can be appropriate cues for detecting tension in this communication context. Our CS/CM annotated interview corpus is openly accessible. Aside from annotating and examining CS/CM occurrences, we annotated silence situations in this open corpus. Silence is shown to be an indicator of tension in interpersonal communications. Making this corpus openly accessible, we call for more research endeavors on tension detection.
Fine-tuning general-purpose pre-trained models has become a de-facto standard, also for Vision and Language tasks such as Visual Question Answering (VQA). In this paper, we take a step back and ask whether a fine-tuned model has superior linguistic and reasoning capabilities than a prior state-of-the-art architecture trained from scratch on the training data alone. We perform a fine-grained evaluation on out-of-distribution data, including an analysis on robustness due to linguistic variation (rephrasings). Our empirical results confirm the benefit of pre-training on overall performance and rephrasing in particular. But our results also uncover surprising limitations, particularly for answering questions involving boolean operations. To complement the empirical evaluation, this paper also surveys relevant earlier work on 1) available VQA data sets, 2) models developed for VQA, 3) pre-trained Vision+Language models, and 4) earlier fine-grained evaluation of pre-trained Vision+Language models.
One of the processing tasks for large multimodal data streams is automatic image description (image classification, object segmentation and classification). Although the number and the diversity of image datasets is constantly expanding, still there is a huge demand for more datasets in terms of variety of domains and object classes covered. The goal of the project Multilingual Image Corpus (MIC 21) is to provide a large image dataset with annotated objects and object descriptions in 24 languages. The Multilingual Image Corpus consists of an Ontology of visual objects (based on WordNet) and a collection of thematically related images whose objects are annotated with segmentation masks and labels describing the ontology classes. The dataset is designed both for image classification and object detection and for semantic segmentation. The main contributions of our work are: a) the provision of large collection of high quality copyright-free images; b) the formulation of the Ontology of visual objects based on WordNet noun hierarchies; c) the precise manual correction of automatic object segmentation within the images and the annotation of object classes; and d) the association of objects and images with extended multilingual descriptions based on WordNet inner- and interlingual relations. The dataset can be used also for multilingual image caption generation, image-to-text alignment and automatic question answering for images and videos.
Sign language production (SLP) is the process of generating sign language videos from spoken language expressions. Since sign languages are highly under-resourced, existing vision-based SLP approaches suffer from out-of-vocabulary (OOV) and test-time generalization problems and thus generate low-quality translations. To address these problems, we introduce an avatar-based SLP system composed of a sign language translation (SLT) model and an avatar animation generation module. Our Transformer-based SLT model utilizes two additional strategies to resolve these problems: named entity transformation to reduce OOV tokens and context vector generation using a pretrained language model (e.g., BERT) to reliably train the decoder. Our system is validated on a new Korean-Korean Sign Language (KSL) dataset of weather forecasts and emergency announcements. Our SLT model achieves an 8.77 higher BLEU-4 score and a 4.57 higher ROUGE-L score over those of our baseline model. In a user evaluation, 93.48% of named entities were successfully identified by participants, demonstrating marked improvement on OOV issues.
We present a five-year retrospective on the development of the VoxWorld platform, first introduced as a multimodal platform for modeling motion language, that has evolved into a platform for rapidly building and deploying embodied agents with contextual and situational awareness, capable of interacting with humans in multiple modalities, and exploring their environments. In particular, we discuss the evolution from the theoretical underpinnings of the VoxML modeling language to a platform that accommodates both neural and symbolic inputs to build agents capable of multimodal interaction and hybrid reasoning. We focus on three distinct agent implementations and the functionality needed to accommodate all of them: Diana, a virtual collaborative agent; Kirby, a mobile robot; and BabyBAW, an agent who self-guides its own exploration of the world.
Posting and sharing memes have become a powerful expedient of expressing opinions on social media in recent days. Analysis of sentiment from memes has gained much attention to researchers due to its substantial implications in various domains like finance and politics. Past studies on sentiment analysis of memes have primarily been conducted in English, where low-resource languages gain little or no attention. However, due to the proliferation of social media usage in recent years, sentiment analysis of memes is also a crucial research issue in low resource languages. The scarcity of benchmark datasets is a significant barrier to performing multimodal sentiment analysis research in resource-constrained languages like Bengali. This paper presents a novel multimodal dataset (named MemoSen) for Bengali containing 4417 memes with three annotated labels positive, negative, and neutral. A detailed annotation guideline is provided to facilitate further resource development in this domain. Additionally, a set of experiments are carried out on MemoSen by constructing twelve unimodal (i.e., visual, textual) and ten multimodal (image+text) models. The evaluation exhibits that the integration of multimodal information significantly improves (about 1.2%) the meme sentiment classification compared to the unimodal counterparts and thus elucidate the novel aspects of multimodality.
We present a new audio-visual speech corpus (RUSAVIC) recorded in a car environment and designed for noise-robust speech recognition. Our goal was to produce a speech corpus which is natural (recorded in real driving conditions), controlled (providing different SNR levels by windows open/closed, moving/parked vehicle, etc.), and adequate size (the amount of data is enough to train state-of-the-art NN approaches). We focus on the problem of audio-visual speech recognition: with the use of automated lip-reading to improve the performance of audio-based speech recognition in the presence of severe acoustic noise caused by road traffic. We also describe the equipment and procedures used to create RUSAVIC corpus. Data are collected in a synchronous way through several smartphones located at different angles and equipped with FullHD video camera and microphone. The corpus includes the recordings of 20 drivers with minimum of 10 recording sessions for each. Besides providing a detailed description of the dataset and its collection pipeline, we evaluate several popular audio and visual speech recognition methods and present a set of baseline recognition results. At the moment RUSAVIC is a unique audio-visual corpus for the Russian language that is recorded in-the-wild condition and we make it publicly available.
This paper presents a corpus of AZee discourse expressions, i.e. expressions which formally describe Sign Language utterances of any length using the AZee approach and language. The construction of this corpus had two main goals: a first reference corpus for AZee, and a test of its coverage on a significant sample of real-life utterances. We worked on productions from an existing corpus, namely the “40 breves”, containing an hour of French Sign Language. We wrote the corresponding AZee discourse expressions for the entire video content, i.e. expressions capturing the forms produced by the signers and their associated meaning by combining known production rules, a basic building block for these expressions. These are made available as a version 2 extension of the “40 breves”. We explain the way in which these expressions can be built, present the resulting corpus and set of production rules used, and perform first measurements on it. We also propose an evaluation of our corpus: for one hour of discourse, AZee allows to describe 94% of it, while ongoing studies are increasing this coverage. This corpus offers a lot of future prospects, for instance concerning synthesis with virtual signers, machine translation or formal grammars for Sign Language.
Evaluating video captioning systems is a challenging task as there are multiple factors to consider; for instance: the fluency of the caption, multiple actions happening in a single scene, and the human bias of what is considered important. Most metrics try to measure how similar the system generated captions are to a single or a set of human-annotated captions. This paper presents a new method based on a deep learning model to evaluate these systems. The model is based on BERT, which is a language model that has been shown to work well in multiple NLP tasks. The aim is for the model to learn to perform an evaluation similar to that of a human. To do so, we use a dataset that contains human evaluations of system generated captions. The dataset consists of the human judgments of the captions produces by the system participating in various years of the TRECVid video to text task. BERTHA obtain favourable results, outperforming the commonly used metrics in some setups.
This paper presents Gesture AMR, an extension to Abstract Meaning Representation (AMR), that captures the meaning of gesture. In developing Gesture AMR, we consider how gesture form and meaning relate; how gesture packages meaning both independently and in interaction with speech; and how the meaning of gesture is temporally and contextually determined. Our case study for developing Gesture AMR is a focused human-human shared task to build block structures. We develop an initial taxonomy of gesture act relations that adheres to AMR’s existing focus on predicate-argument structure while integrating meaningful elements unique to gesture. Pilot annotation shows Gesture AMR to be more challenging than standard AMR, and illustrates the need for more work on representation of dialogue and multimodal meaning. We discuss challenges of adapting an existing meaning representation to non-speech-based modalities and outline several avenues for expanding Gesture AMR.
This paper presents a new training dataset for automatic genre identification GINCO, which is based on 1,125 crawled Slovenian web documents that consist of 650,000 words. Each document was manually annotated for genre with a new annotation schema that builds upon existing schemata, having primarily clarity of labels and inter-annotator agreement in mind. The dataset consists of various challenges related to web-based data, such as machine translated content, encoding errors, multiple contents presented in one document etc., enabling evaluation of classifiers in realistic conditions. The initial machine learning experiments on the dataset show that (1) pre-Transformer models are drastically less able to model the phenomena, with macro F1 metrics ranging around 0.22, while Transformer-based models achieve scores of around 0.58, and (2) multilingual Transformer models work as well on the task as the monolingual models that were previously proven to be superior to multilingual models on standard NLP tasks.
With the emergence of neural end-to-end approaches for spoken language understanding (SLU), a growing number of studies have been presented during these last three years on this topic. The major part of these works addresses the spoken language understanding domain through a simple task like speech intent detection. In this context, new benchmark datasets have also been produced and shared with the community related to this task. In this paper, we focus on the French MEDIA SLU dataset, distributed since 2005 and used as a benchmark dataset for a large number of research works. This dataset has been shown as being the most challenging one among those accessible to the research community. Distributed by ELRA, this corpus is free for academic research since 2019. Unfortunately, the MEDIA dataset is not really used beyond the French research community. To facilitate its use, a complete recipe, including data preparation, training and evaluation scripts, has been built and integrated to SpeechBrain, an already popular open-source and all-in-one conversational AI toolkit based on PyTorch. This recipe is presented in this paper. In addition, based on the feedback of some researchers who have worked on this dataset for several years, some corrections have been brought to the initial manual annotation: the new version of the data will also be integrated into the ELRA catalogue, as the original one. More, a significant amount of data collected during the construction of the MEDIA corpus in the 2000s was never used until now: we present the first results reached on this subset — also included in the MEDIA SpeechBrain recipe — , that will be used for now as the MEDIA test2. Last, we discuss evaluation issues.
Natural Language Understanding (NLU) technology has improved significantly over the last few years and multitask benchmarks such as GLUE are key to evaluate this improvement in a robust and general way. These benchmarks take into account a wide and diverse set of NLU tasks that require some form of language understanding, beyond the detection of superficial, textual clues. However, they are costly to develop and language-dependent, and therefore they are only available for a small number of languages. In this paper, we present BasqueGLUE, the first NLU benchmark for Basque, a less-resourced language, which has been elaborated from previously existing datasets and following similar criteria to those used for the construction of GLUE and SuperGLUE. We also report the evaluation of two state-of-the-art language models for Basque on BasqueGLUE, thus providing a strong baseline to compare upon. BasqueGLUE is freely available under an open license.
This paper presents, to the best of our knowledge, the first ever publicly available annotated dataset for sentiment classification and semantic polarity dictionary for Georgian. The characteristics of these resources and the process of their creation are described in detail. The results of various experiments on the performance of both lexicon- and machine learning-based models for Georgian sentiment classification are also reported. Both 3-label (positive, neutral, negative) and 4-label settings (same labels + mixed) are considered. The machine learning models explored include, i.a., logistic regression, SVMs, and transformed-based models. We also explore transfer learning- and translation-based (to a well-supported language) approaches. The obtained results for Georgian are on par with the state-of-the-art results in sentiment classification for well studied languages when using training data of comparable size.
Natural Language Processing is increasingly being applied in the finance and business industry to analyse the text of many different types of financial documents. Given the increasing growth of firms around the world, the volume of financial disclosures and financial texts in different languages and forms is increasing sharply and therefore the study of language technology methods that automatically summarise content has grown rapidly into a major research area. Corpora for financial narrative summarisation exists in English, but there is a significant lack of financial text resources in the French language. To remedy this, we present CoFiF Plus, the first French financial narrative summarisation dataset providing a comprehensive set of financial text written in French. The dataset has been extracted from French financial reports published in PDF file format. It is composed of 1,703 reports from the most capitalised companies in France (Euronext Paris) covering a time frame from 1995 to 2021. This paper describes the collection, annotation and validation of the financial reports and their summaries. It also describes the dataset and gives the results of some baseline summarisers. Our datasets will be openly available upon the acceptance of the paper.
Almost all summarisation methods and datasets focus on a single language and short summaries. We introduce a new dataset called WikinewsSum for English, German, French, Spanish, Portuguese, Polish, and Italian summarisation tailored for extended summaries of approx. 11 sentences. The dataset comprises 39,626 summaries which are news articles from Wikinews and their sources. We compare three multilingual transformer models on the extractive summarisation task and three training scenarios on which we fine-tune mT5 to perform abstractive summarisation. This results in strong baselines for both extractive and abstractive summarisation on WikinewsSum. We also show how the combination of an extractive model with an abstractive one can be used to create extended abstractive summaries from long input documents. Finally, our results show that fine-tuning mT5 on all the languages combined significantly improves the summarisation performance on low-resource languages.
Progress in sentence simplification has been hindered by a lack of labeled parallel simplification data, particularly in languages other than English. We introduce MUSS, a Multilingual Unsupervised Sentence Simplification system that does not require labeled simplification data. MUSS uses a novel approach to sentence simplification that trains strong models using sentence-level paraphrase data instead of proper simplification data. These models leverage unsupervised pretraining and controllable generation mechanisms to flexibly adjust attributes such as length and lexical complexity at inference time. We further present a method to mine such paraphrase data in any language from Common Crawl using semantic sentence embeddings, thus removing the need for labeled data. We evaluate our approach on English, French, and Spanish simplification benchmarks and closely match or outperform the previous best supervised results, despite not using any labeled simplification data. We push the state of the art further by incorporating labeled simplification data.
Women are often perceived as junior to their male counterparts, even within the same job titles. While there has been significant progress in the evaluation of gender bias in natural language processing (NLP), existing studies seldom investigate how biases toward gender groups change when compounded with other societal biases. In this work, we investigate how seniority impacts the degree of gender bias exhibited in pretrained neural generation models by introducing a novel framework for probing compound bias. We contribute a benchmark robustness-testing dataset spanning two domains, U.S. senatorship and professorship, created using a distant-supervision method. Our dataset includes human-written text with underlying ground truth and paired counterfactuals. We then examine GPT-2 perplexity and the frequency of gendered language in generated text. Our results show that GPT-2 amplifies bias by considering women as junior and men as senior more often than the ground truth in both domains. These results suggest that NLP applications built using GPT-2 may harm women in professional capacities.
This paper presents contributions in two directions: first we propose a new system for Frame Identification (FI), based on pre-trained text encoders trained discriminatively and graphs embedding, producing state of the art performance and, second, we take in consideration all the extremely different procedures used to evaluate systems for this task performing a complete evaluation over two benchmarks and all possible splits and cleaning procedures used in the FI literature.
Our discourses are full of potential lexical ambiguities, due in part to the pervasive use of words having multiple senses. Sometimes, one word may even be used in more than one sense throughout a text. But, to what extent is this true for different kinds of texts? Does the use of polysemous words change when a discourse involves two people, or when speakers have time to plan what to say? We investigate these questions by comparing the polysemy level of texts of different nature, with a focus on spontaneous spoken dialogs; unlike previous work which examines solely scripted, written, monolog-like data. We compare multiple metrics that presuppose different conceptualizations of text polysemy, i.e., they consider the observed or the potential number of senses of words, or their sense distribution in a discourse. We show that the polysemy level of texts varies greatly depending on the kind of text considered, with dialog and spoken discourses having generally a higher polysemy level than written monologs. Additionally, our results emphasize the need for relaxing the popular “one sense per discourse” hypothesis.
Cross-Level Semantic Similarity (CLSS) is a measure of the level of semantic overlap between texts of different lengths. Although this problem was formulated almost a decade ago, research on it has been sparse, and limited exclusively to the English language. In this paper, we present the first CLSS dataset in another language, in the form of CLSS.news.sr – a corpus of 1000 phrase-sentence and 1000 sentence-paragraph newswire text pairs in Serbian, manually annotated with fine-grained semantic similarity scores using a 0–4 similarity scale. We describe the methodology of data collection and annotation, and compare the resulting corpus to its preexisting counterpart in English, SemEval CLSS, following up with a preliminary linguistic analysis of the newly created dataset. State-of-the-art pre-trained language models are then fine-tuned and evaluated on the CLSS task in Serbian using the produced data, and their settings and results are discussed. The CLSS.news.sr corpus and the guidelines used in its creation are made publicly available.
Semantic role labeling (SRL) represents the meaning of a sentence in the form of predicate-argument structures. Such shallow semantic analysis is helpful in a wide range of downstream NLP tasks and real-world applications. As treebanks enabled the development of powerful syntactic parsers, the accurate predicate-argument analysis demands training data in the form of propbanks. Unfortunately, most languages simply do not have corresponding propbanks due to the high cost required to construct such resources. To overcome such challenges, Universal Proposition Bank 1.0 (UP1.0) was released in 2017, with high-quality propbank data generated via a two-stage method exploiting monolingual SRL and multilingual parallel data. In this paper, we introduce Universal Proposition Bank 2.0 (UP2.0), with significant enhancements over UP1.0: (1) propbanks with higher quality by using a state-of-the-art monolingual SRL and improved auto-generation of annotations; (2) expanded language coverage (from 7 to 9 languages); (3) span annotation for the decoupling of syntactic analysis; and (4) Gold data for a subset of the languages. We also share our experimental results that confirm the significant quality improvements of the generated propbanks. In addition, we present a comprehensive experimental evaluation on how different implementation choices impact the quality of the resulting data. We release these resources to the research community and hope to encourage more research on cross-lingual SRL.
Eye movement recordings from reading are one of the richest signals of human language processing. Corpora of eye movements during reading of contextualized running text is a way of making such records available for natural language processing purposes. Such corpora already exist in some languages. We present CopCo, the Copenhagen Corpus of eye tracking recordings from natural reading of Danish texts. It is the first eye tracking corpus of its kind for the Danish language. CopCo includes 1,832 sentences with 34,897 tokens of Danish text extracted from a collection of speech manuscripts. This first release of the corpus contains eye tracking data from 22 participants. It will be extended continuously with more participants and texts from other genres. We assess the data quality of the recorded eye movements and find that the extracted features are in line with related research. The dataset available here: https://osf.io/ud8s5/.
We present the Brooklyn Multi-Interaction Corpus (B-MIC), a collection of dyadic conversations designed to identify speaker traits and conversation contexts that cause variations in entrainment behavior. B-MIC pairs each participant with multiple partners for an object placement game and open-ended discussions, as well as with a Wizard of Oz for a baseline of their speech. In addition to fully transcribed recordings, it includes demographic information and four completed psychological questionnaires for each subject and turn annotations for perceived emotion and acoustic outliers. This enables the study of speakers’ entrainment behavior in different contexts and the sources of variation in this behavior. In this paper, we introduce B-MIC and describe our collection, annotation, and preprocessing methodologies. We report a preliminary study demonstrating varied entrainment behavior across different conversation types and discuss the rich potential for future work on the corpus.
Pro-TEXT is a corpus of keystroke logs written in French. Keystroke logs are recordings of the writing process executed through a keyboard, which keep track of all actions taken by the writer (character additions, deletions, substitutions). As such, the Pro-TEXT corpus offers new insights into text genesis and underlying cognitive processes from the production perspective. A subset of the corpus is linguistically annotated with parts of speech, lemmas and syntactic dependencies, making it suitable for the study of interactions between linguistic and behavioural aspects of the writing process. The full corpus contains 202K tokens, while the annotated portion is currently 30K tokens large. The annotated content is progressively being made available in a database-like CSV format and in CoNLL format, and the work on an HTML-based visualisation tool is currently under way. To the best of our knowledge, Pro-TEXT is the first corpus of its kind in French.
Corpus-based studies on acceptability judgements have always stimulated the interest of researchers, both in theoretical and computational fields. Some approaches focused on spontaneous judgements collected through different types of tasks, others on data annotated through crowd-sourcing platforms, still others relied on expert annotated data available from the literature. The release of CoLA corpus, a large-scale corpus of sentences extracted from linguistic handbooks as examples of acceptable/non acceptable phenomena in English, has revived interest in the reliability of judgements of linguistic experts vs. non-experts. Several issues are still open. In this work, we contribute to this debate by presenting a 3D video game that was used to collect acceptability judgments on Italian sentences. We analyse the resulting annotations in terms of agreement among players and by comparing them with experts’ acceptability judgments. We also discuss different game settings to assess their impact on participants’ motivation and engagement. The final dataset containing 1,062 sentences, which were selected based on majority voting, is released for future research and comparisons.
This paper describes a new corpus of human translations which contains both professional and students translations. The data consists of English sources – texts from news and reviews – and their translations into Russian and Croatian, as well as of the subcorpus containing translations of the review texts into Finnish. All target languages represent mid-resourced and less or mid-investigated ones. The corpus will be valuable for studying variation in translation as it allows a direct comparison between human translations of the same source texts. The corpus will also be a valuable resource for evaluating machine translation systems. We believe that this resource will facilitate understanding and improvement of the quality issues in both human and machine translation. In the paper, we describe how the data was collected, provide information on translator groups and summarise the differences between the human translations at hand based on our preliminary results with shallow features.
Automatic identification of cyberbullying from textual content is known to be a challenging task. The challenges arise from the inherent structure of cyberbullying and the lack of labeled large-scale corpus, enabling efficient machine-learning-based tools including neural networks. This paper advocates a data augmentation-based approach that could enhance the automatic detection of cyberbullying in social media texts. We use both word sense disambiguation and synonymy relation in WordNet lexical database to generate coherent equivalent utterances of cyberbullying input data. The disambiguation and semantic expansion are intended to overcome the inherent limitations of social media posts, such as an abundance of unstructured constructs and limited semantic content. Besides, to test the feasibility, a novel protocol has been employed to collect cyberbullying traces data from AskFm forum, where about a 10K-size dataset has been manually labeled. Next, the problem of cyberbullying identification is viewed as a binary classification problem using an elaborated data augmentation strategy and an appropriate classifier. For the latter, a Convolutional Neural Network (CNN) architecture with FastText and BERT was put forward, whose results were compared against commonly employed Naïve Bayes (NB) and Logistic Regression (LR) classifiers with and without data augmentation. The research outcomes were promising and yielded almost 98.4% of classifier accuracy, an improvement of more than 4% over baseline results.
Summarization is a challenging problem, and even more challenging is to manually create, correct, and evaluate the summaries. The severity of the problem grows when the inputs are multi-party dialogues in a meeting setup. To facilitate the research in this area, we present ALIGNMEET, a comprehensive tool for meeting annotation, alignment, and evaluation. The tool aims to provide an efficient and clear interface for fast annotation while mitigating the risk of introducing errors. Moreover, we add an evaluation mode that enables a comprehensive quality evaluation of meeting minutes. To the best of our knowledge, there is no such tool available. We release the tool as open source. It is also directly installable from PyPI.
Stuttering is a complex speech disorder that negatively affects an individual’s ability to communicate effectively. Persons who stutter (PWS) often suffer considerably under the condition and seek help through therapy. Fluency shaping is a therapy approach where PWSs learn to modify their speech to help them to overcome their stutter. Mastering such speech techniques takes time and practice, even after therapy. Shortly after therapy, success is evaluated highly, but relapse rates are high. To be able to monitor speech behavior over a long time, the ability to detect stuttering events and modifications in speech could help PWSs and speech pathologists to track the level of fluency. Monitoring could create the ability to intervene early by detecting lapses in fluency. To the best of our knowledge, no public dataset is available that contains speech from people who underwent stuttering therapy that changed the style of speaking. This work introduces the Kassel State of Fluency (KSoF), a therapy-based dataset containing over 5500 clips of PWSs. The clips were labeled with six stuttering-related event types: blocks, prolongations, sound repetitions, word repetitions, interjections, and – specific to therapy – speech modifications. The audio was recorded during therapy sessions at the Institut der Kasseler Stottertherapie. The data will be made available for research purposes upon request.
Users generate content constantly, leading to new data requiring annotation. Among this data, textual conversations are created every day and come with some specificities: they are mostly private through instant messaging applications, requiring the conversational context to be labeled. These specificities led to several annotation tools dedicated to conversation, and mostly dedicated to dialogue tasks, requiring complex annotation schemata, not always customizable and not taking into account conversation-level labels. In this paper, we present EZCAT, an easy-to-use interface to annotate conversations in a two-level configurable schema, leveraging message-level labels and conversation-level labels. Our interface is characterized by the voluntary absence of a server and accounts management, enhancing its availability to anyone, and the control over data, which is crucial to confidential conversations. We also present our first usage of EZCAT along with our annotation schema we used to annotate confidential customer service conversations. EZCAT is freely available at https://gguibon.github.io/ezcat.
Given the benefits of syntactically annotated collections of transcribed speech in spoken language research and applications, many spoken language treebanks have been developed in the last decades, with divergent annotation schemes posing important limitations to cross-resource explorations, such as comparing data across languages, grammatical frameworks, and language domains. As a consequence, there has been a growing number of spoken language treebanks adopting the Universal Dependencies (UD) annotation scheme, aimed at cross-linguistically consistent morphosyntactic annotation. In view of the non-central role of spoken language data within the scheme and with little in-domain consolidation to date, this paper presents a comparative overview of spoken language treebanks in UD to support cross-treebank data explorations on the one hand, and encourage further treebank harmonization on the other. Our results show that the spoken language treebanks differ considerably with respect to the inventory and the format of transcribed phenomena, as well as the principles adopted in their morphosyntactic annotation. This is particularly true for the dependency annotation of speech disfluencies, where conflicting data annotations suggest an underspecification of the guidelines pertaining to speech repairs in general and the reparandum dependency relation in particular.
We present LeConTra, a learner corpus consisting of English-to-Dutch news translations enriched with translation process data. Three students of a Master’s programme in Translation were asked to translate 50 different English journalistic texts of approximately 250 tokens each. Because we also collected translation process data in the form of keystroke logging, our dataset can be used as part of different research strands such as translation process research, learner corpus research, and corpus-based translation studies. Reference translations, without process data, are also included. The data has been manually segmented and tokenized, and manually aligned at both segment and word level, leading to a high-quality corpus with token-level process data. The data is freely accessible via the Translation Process Research DataBase, which emphasises our commitment of distributing our dataset. The tool that was built for manual sentence segmentation and tokenization, Mantis, is also available as an open-source aid for data processing.
This paper focuses on detection of sources in the Czech articles published on a news server of Czech public radio. In particular, we search for attribution in sentences and we recognize attributed sources and their sentence context (signals). We organized a crowdsourcing annotation task that resulted in a data set of 2,167 stories with manually recognized signals and sources. In addition, the sources were classified into the classes of named and unnamed sources.
We present Xposition, an online platform for documenting adpositional semantics across languages in terms of supersenses (Schneider et al., 2018). More than just a lexical database, Xposition houses annotation guidelines, structured lexicographic documentation, and annotated corpora. Guidelines and documentation are stored as wiki pages for ease of editing, and described elements (supersenses, adpositions, etc.) are hyperlinked for ease of browsing. We describe how the platform structures information; its current contents across several languages; and aspects of the design of the web application that supports it, with special attention to how it supports datasets and standards that evolve over time.
This paper describes data resources created for Phase 1 of the DARPA Active Interpretation of Disparate Alternatives (AIDA) program, which aims to develop language technology that can help humans manage large volumes of sometimes conflicting information to develop a comprehensive understanding of events around the world, even when such events are described in multiple media and languages. Especially important is the need for the technology to be capable of building multiple hypotheses to account for alternative interpretations of data imbued with informational conflict. The corpus described here is designed to support these goals. It focuses on the domain of Russia-Ukraine relations and contains multimedia source data in English, Russian and Ukrainian, annotated to support development and evaluation of systems that perform extraction of entities, events, and relations from individual multimedia documents, aggregate the information across documents and languages, and produce multiple “hypotheses” about what has happened. This paper describes source data collection, annotation, and assessment.
This paper describes our efforts to extend the PARSEME framework to Modern Standard Arabic. Theapplicability of the PARSEME guidelines was tested by measuring the inter-annotator agreement in theearly annotation stage. A subset of 1,062 sentences from the Prague Arabic Dependency Treebank PADTwas selected and annotated by two Arabic native speakers independently. Following their annotations, anew Arabic corpus with over 1,250 annotated VMWEs has been built. This corpus already exceeds thesmallest corpora of the PARSEME suite, and enables first observations. We discuss our annotation guide-line schema that shows full MWE annotation is realizable in Arabic where we get good inter-annotator agreement.
There has been a lot of work on predicting the timing of feedback in conversational systems. However, there has been less focus on predicting the prosody and lexical form of feedback given their communicative function. Therefore, in this paper we present our preliminary annotations of the communicative functions of 1627 short feedback tokens from the Switchboard corpus and an analysis of their lexical realizations and prosodic characteristics. Since there is no standard scheme for annotating the communicative function of feedback we propose our own annotation scheme. Although our work is ongoing, our preliminary analysis revealed lexical tokens such as “yeah” are ambiguous and therefore lexical forms alone are not indicative of the function. Both the lexical form and prosodic characteristics need to be taken into account in order to predict the communicative function. We also found that feedback functions have distinguishable prosodic characteristics in terms of duration, mean pitch, pitch slope, and pitch range.
Social media are a central part of people’s lives. Unfortunately, many public social media spaces are rife with bullying and offensive language, creating an unsafe environment for their users. In this paper, we present a new dataset for offensive language detection in Albanian. The dataset is composed of user-generated comments on Facebook and YouTube from the channels of selected Kosovo news platforms. It is annotated according to the three levels of the OLID annotation scheme. We also show results of a baseline system for offensive language classification based on a fine-tuned BERT model and compare with the Danish DKhate dataset, which is similar in scope and size. In a transfer learning setting, we find that merging the Albanian and Danish training sets leads to improved performance for prediction on Danish, but not Albanian, on both offensive language recognition and distinguishing targeted and untargeted offence.
Gender bias in natural language processing (NLP) applications, particularly machine translation, has been receiving increasing attention. Much of the research on this issue has focused on mitigating gender bias in English NLP models and systems. Addressing the problem in poorly resourced, and/or morphologically rich languages has lagged behind, largely due to the lack of datasets and resources. In this paper, we introduce a new corpus for gender identification and rewriting in contexts involving one or two target users (I and/or You) – first and second grammatical persons with independent grammatical gender preferences. We focus on Arabic, a gender-marking morphologically rich language. The corpus has multiple parallel components: four combinations of 1st and 2nd person in feminine and masculine grammatical genders, as well as English, and English to Arabic machine translation output. This corpus expands on Habash et al. (2019)’s Arabic Parallel Gender Corpus (APGC v1.0) by adding second person targets as well as increasing the total number of sentences over 6.5 times, reaching over 590K words. Our new dataset will aid the research and development of gender identification, controlled text generation, and post-editing rewrite systems that could be used to personalize NLP applications and provide users with the correct outputs based on their grammatical gender preferences. We make the Arabic Parallel Gender Corpus (APGC v2.0) publicly available
Social media platforms play an increasingly important role as forums for public discourse. Many platforms use recommendation algorithms that funnel users to online groups with the goal of maximizing user engagement, which many commentators have pointed to as a source of polarization and misinformation. Understanding the role of NLP in recommender systems is an interesting research area, given the role that social media has played in world events. However, there are few standardized resources which researchers can use to build models that predict engagement with online groups on social media; each research group constructs datasets from scratch without releasing their version for reuse. In this work, we present a dataset drawn from posts and comments on the online message board Reddit. We develop baseline models for recommending subreddits to users, given the user’s post and comment history. We also study the behavior of our recommender models on subreddits that were banned in June 2020 as part of Reddit’s efforts to stop the dissemination of hate speech.
Interest in argument mining has resulted in an increasing number of argument annotated corpora. However, most focus on English texts with explicit argumentative discourse markers, such as persuasive essays or legal documents. Conversely, we report on the first extensive and consolidated Portuguese argument annotation project focused on opinion articles. We briefly describe the annotation guidelines based on a multi-layered process and analyze the manual annotations produced, highlighting the main challenges of this textual genre. We then conduct a comprehensive inter-annotator agreement analysis, including argumentative discourse units, their classes and relations, and resulting graphs. This analysis reveals that each of these aspects tackles very different kinds of challenges. We observe differences in annotator profiles, motivating our aim of producing a non-aggregated corpus containing the insights of every annotator. We note that the interpretation and identification of token-level arguments is challenging; nevertheless, tasks that focus on higher-level components of the argument structure can obtain considerable agreement. We lay down perspectives on corpus usage, exploiting its multi-faceted nature.
Parliamentary debates represent a large and partly unexploited treasure trove of publicly accessible texts. In the German-speaking area, there is a certain deficit of uniformly accessible and annotated corpora covering all German-speaking parliaments at the national and federal level. To address this gap, we introduce the German Parliamentary Corpus (GerParCor). GerParCor is a genre-specific corpus of (predominantly historical) German-language parliamentary protocols from three centuries and four countries, including state and federal level data. In addition, GerParCor contains conversions of scanned protocols and, in particular, of protocols in Fraktur converted via an OCR process based on Tesseract. All protocols were preprocessed by means of the NLP pipeline of spaCy3 and automatically annotated with metadata regarding their session date. GerParCor is made available in the XMI format of the UIMA project. In this way, GerParCor can be used as a large corpus of historical texts in the field of political communication for various tasks in NLP.
In this paper, we present an upgraded version of the Hungarian NYTK-NerKor named entity corpus, which contains about twice as many annotated spans and 7 times as many distinct entity types as the original version. We used an extended version of the OntoNotes 5 annotation scheme including time and numerical expressions. NerKor is the newest and biggest NER corpus for Hungarian containing diverse domains. We applied cross-lingual transfer of NER models trained for other languages based on multilingual contextual language models to preannotate the corpus. We corrected the annotation semi-automatically and manually. Zero-shot preannotation was very effective with about 0.82 F1 score for the best model. We also added a 12000-token subcorpus on cars and other motor vehicles. We trained and release a transformer-based NER tagger for Hungarian using the annotation in the new corpus version, which provides similar performance to an identical model trained on the original version of the corpus.
Since several decades emotional databases have been recorded by various laboratories. Many of them contain acted portrays of Darwin’s famous “big four” basic emotions. In this paper, we investigate in how far a selection of them are comparable by two approaches: on the one hand modeling similarity as performance in cross database machine learning experiments and on the other by analyzing a manually picked set of four acoustic features that represent different phonetic areas. It is interesting to see in how far specific databases (we added a synthetic one) perform well as a training set for others while some do not. Generally speaking, we found indications for both similarity as well as specificiality across languages.
We present advancements with a software tool called Nkululeko, that lets users perform (semi-) supervised machine learning experiments in the speaker characteristics domain. It is based on audformat, a format for speech database metadata description. Due to an interface based on configurable templates, it supports best practise and very fast setup of experiments without the need to be proficient in the underlying language: Python. The paper explains the handling of Nkululeko and presents two typical experiments: comparing the expert acoustic features with artificial neural net embeddings for emotion classification and speaker age regression.
Aerodynamic processes underlie the characteristics of the acoustic signal of speech sounds. The aerodynamics of speech give insights on acoustic outcome and help explain the mechanisms of speech production. This database was designed during an ARC project ”Dynamique des systèmes phonologiques” in which the study of aerodynamic constraints on speech production was an important target. Data were recorded between 1996 and 1999 at the Erasmus Hospital (Hôpital Erasme) of Université Libre de Bruxelles, Belgium and constitute one of the few datasets available on direct measurement of subglottal pressure and other aerodynamic parameters. The goal was to obtain a substantial amount of data with simultaneous recording, in various context, of the speech acoustic signal, subglottal pressure (Ps), intraoral pressure (Po), oral airflow (Qo) and nasal airflow (Qn). This database contains recordings of 2 English, 1 Amharic, and 7 French speakers and is provided with data conversion and visualisation tools. Another aim of this project was to obtain some reference values of the aerodynamics of speech production for female and male speakers uttering different types of segments and sentences in French.
Our knowledge on speech is historically built on data comparing different speakers or data averaged across speakers. Consequently, little is known on the variability in the speech of a single individual. Experimental studies have shown that speakers adapt to the linguistic and the speaking contexts, and modify their speech according to their emotional or biological condition, etc. However, it is unclear how much speakers vary from one repetition to the next, and how comparable are recordings that are collected days, months or years apart. In this paper, we introduce two French databases which contain recordings of 9 to 11 speakers recorded over 9 to 18 sessions, allowing comparisons of speech tasks with a different delay between the repetitions: 3 repetitions within the same session, 6 to 10 repetitions on different days during a two months period, 5 to 9 repetitions on different years. Speakers are recorded on a large set of speech tasks including read and spontaneous speech as well as speech-like performance tasks. In this paper, we provide detailed descriptions of the two databases and available annotations. We conclude by an illustration on how these data can inform on within-speaker variability of speech.
Building a usable radio monitoring automatic speech recognition (ASR) system is a challenging task for under-resourced languages and yet this is paramount in societies where radio is the main medium of public communication and discussions. Initial efforts by the United Nations in Uganda have proved how understanding the perceptions of rural people who are excluded from social media is important in national planning. However, these efforts are being challenged by the absence of transcribed speech datasets. In this paper, The Makerere Artificial Intelligence research lab releases a Luganda radio speech corpus of 155 hours. To our knowledge, this is the first publicly available radio dataset in sub-Saharan Africa. The paper describes the development of the voice corpus and presents baseline Luganda ASR performance results using Coqui STT toolkit, an open-source speech recognition toolkit.
In this paper, we present a far-field speaker verification benchmark derived from the publicly-available DiPCo corpus. This corpus comprise three different tasks that involve enrollment and test conditions with single- and/or multi-channels recordings. The main goal of this corpus is to foster research in far-field and multi-channel text-independent speaker verification. Also, it can be used for other speaker recognition tasks such as dereverberation, denoising and speech enhancement. In addition, we release a Kaldi and SpeechBrain system to facilitate further research. And we validate the evaluation design with a single-microphone state-of-the-art speaker recognition system (i.e. ResNet-101). The results show that the proposed tasks are very challenging. And we hope these resources will inspire the speech community to develop new methods and systems for this challenging domain.
Inserting fillers (such as “um”, “like”) to clean speech text has a rich history of study. One major application is to make dialogue systems sound more spontaneous. The ambiguity of filler occurrence and inter-speaker difference make both modeling and evaluation difficult. In this paper, we study sampling-based filler insertion, a simple yet unexplored approach to inserting fillers. We propose an objective score called Filler Perplexity (FPP). We build three models trained on two single-speaker spontaneous corpora, and evaluate them with FPP and perceptual tests. We implement two innovations in perceptual tests, (1) evaluating filler insertion on dialogue systems output, (2) synthesizing speech with neural spontaneous TTS engines. FPP proves to be useful in analysis but does not correlate well with perceptual MOS. Perceptual results show little difference between compared filler insertion models including with ground-truth, which may be due to the ambiguity of what is good filler insertion and a strong neural spontaneous TTS that produces natural speech irrespective of input. Results also show preference for filler-inserted speech synthesized with spontaneous TTS. The same test using TTS based on read speech obtains the opposite results, which shows the importance of using spontaneous TTS in evaluating filler insertions. Audio samples: www.speech.kth.se/tts-demos/LREC22
Hungarian is spoken by 15 million people, still, easily accessible Automatic Speech Recognition (ASR) benchmark datasets – especially for spontaneous speech – have been practically unavailable. In this paper, we introduce BEA-Base, a subset of the BEA spoken Hungarian database comprising mostly spontaneous speech of 140 speakers. It is built specifically to assess ASR, primarily for conversational AI applications. After defining the speech recognition subsets and task, several baselines – including classic HMM-DNN hybrid and end-to-end approaches augmented by cross-language transfer learning – are developed using open-source toolkits. The best results obtained are based on multilingual self-supervised pretraining, achieving a 45% recognition error rate reduction as compared to the classical approach – without the application of an external language model or additional supervised data. The results show the feasibility of using BEA-Base for training and evaluation of Hungarian speech recognition systems.
We present SNuC, the first published corpus of spoken alphanumeric identifiers of the sort typically used as serial and part numbers in the manufacturing sector. The dataset contains recordings and transcriptions of over 50 native British English speakers, speaking over 13,000 multi-character alphanumeric sequences and totalling almost 20 hours of recorded speech. We describe requirements taken into account in the designing the corpus and the methodology used to construct it. We present summary statistics describing the corpus contents, as well as a preliminary investigation into errors in spoken alphanumeric identifiers. We validate the corpus by showing how it can be used to adapt a deep learning neural network based ASR system, resulting in improved recognition accuracy on the task of spoken alphanumeric identifier recognition. Finally, we discuss further potential uses for the corpus and for the tools developed to construct it.
In the present paper, we introduce the ManDi Corpus, a spoken corpus of regional Mandarin dialects and Standard Mandarin. The corpus currently contains 357 recordings (about 9.6 hours) of monosyllabic words, disyllabic words, short sentences, a short passage and a poem, each produced in Standard Mandarin and in one of six regional Mandarin dialects: Beijing, Chengdu, Jinan, Taiyuan, Wuhan, and Xi’an Mandarin from 36 speakers. The corpus was collected remotely using participant-controlled smartphone recording apps. Word- and phone-level alignments were generated using Praat and the Montreal Forced Aligner. The pilot study of dialect-specific tone systems showed that with practicable design and decent recording quality, remotely collected speech data can be suitable for analysis of relative patterns in acoustic-phonetic realization. The corpus is available on OSF (https://osf.io/fgv4w/) for non-commercial use under a CC BY-NC 3.0 license.
Conversations (normal speech) or professional interactions (e.g., projected speech in the classroom) have been identified as situations with increased risk of exposure to SARS-CoV-2 due to the high production of droplets in the exhaled air. However, it is still unclear to what extent speech properties influence droplets emission during everyday life conversations. Here, we report the experimental protocol of three experiments aiming at measuring the velocity and the direction of the airflow, the number and size of droplets spread during speech interactions in French. We consider different phonetic conditions, potentially leading to a modulation of speech droplets production, such as voice intensity (normal vs. loud voice), articulation manner of phonemes (type of consonants and vowels) and prosody (i.e., the melody of the speech). Findings from these experiments will allow future simulation studies to predict the transport, dispersion and evaporation of droplets emitted under different speech conditions.
The growing popularity of various forms of Spoken Dialogue Systems (SDS) raises the demand for their capability of implicitly assessing the speaker’s sentiment from speech only. Mapping the latter on user preferences enables to adapt to the user and individualize the requested information while increasing user satisfaction. In this paper, we explore the integration of rank consistent ordinal regression into a speech-only sentiment prediction task performed by ResNet-like systems. Furthermore, we use speaker verification extractors trained on larger datasets as low-level feature extractors. An improvement of performance is shown by fusing sentiment and pre-extracted speaker embeddings reducing the speaker bias of sentiment predictions. Numerous experiments on Multimodal Opinion Sentiment and Emotion Intensity (CMU-MOSEI) databases show that we beat the baselines of state-of-the-art unimodal approaches. Using speech as the only modality combined with optimizing an order-sensitive objective function gets significantly closer to the sentiment analysis results of state-of-the-art multimodal systems.
This research explores automated text classification using data from Low– and Middle–Income Countries (LMICs). In particular, we explore enhancing text representations with demographic information of speakers in a privacy-preserving manner. We introduce the Demographic-Rich Qualitative UPV-Interviews Dataset (DR-QI), a rich dataset of qualitative interviews from rural communities in India and Uganda. The interviews were conducted following the latest standards for respectful interactions with illiterate speakers (Hirmer et al., 2021a). The interviews were later sentence-annotated for Automated User-Perceived Value (UPV) Classification (Conforti et al., 2020), a schema that classifies values expressed by speakers, resulting in a dataset of 5,333 sentences. We perform the UPV classification task, which consists of predicting which values are expressed in a given sentence, on the new DR-QI dataset. We implement a classification model using DistilBERT (Sanh et al., 2019), which we extend with demographic information. In order to preserve the privacy of speakers, we investigate encoding demographic information using autoencoders. We find that adding demographic information improves performance, even if such information is encoded. In addition, we find that the performance per UPV is linked to the number of occurrences of that value in our data.
For research in audiovisual interview archives often it is not only of interest what is said but also how. Sentiment analysis and emotion recognition can help capture, categorize and make these different facets searchable. In particular, for oral history archives, such indexing technologies can be of great interest. These technologies can help understand the role of emotions in historical remembering. However, humans often perceive sentiments and emotions ambiguously and subjectively. Moreover, oral history interviews have multi-layered levels of complex, sometimes contradictory, sometimes very subtle facets of emotions. Therefore, the question arises of the chance machines and humans have capturing and assigning these into predefined categories. This paper investigates the ambiguity in human perception of emotions and sentiment in German oral history interviews and the impact on machine learning systems. Our experiments reveal substantial differences in human perception for different emotions. Furthermore, we report from ongoing machine learning experiments with different modalities. We show that the human perceptual ambiguity and other challenges, such as class imbalance and lack of training data, currently limit the opportunities of these technologies for oral history archives. Nonetheless, our work uncovers promising observations and possibilities for further research.
Finding the polarity of feelings in texts is a far-reaching task. Whilst the field of natural language processing has established sentiment analysis as an alluring problem, many feelings are left uncharted. In this study, we analyze the optimism and pessimism concepts from Twitter posts to effectively understand the broader dimension of psychological phenomenon. Towards this, we carried a systematic study by first exploring the linguistic peculiarities of optimism and pessimism in user-generated content. Later, we devised a multi-task knowledge distillation framework to simultaneously learn the target task of optimism detection with the help of the auxiliary task of sentiment analysis and hate speech detection. We evaluated the performance of our proposed approach on the benchmark Optimism/Pessimism Twitter dataset. Our extensive experiments show the superior- ity of our approach in correctly differentiating between optimistic and pessimistic users. Our human and automatic evaluation shows that sentiment analysis and hate speech detection are beneficial for optimism/pessimism detection.
In this paper, we present a student feedback corpus, which contains 3000 instances of feedback written by university students. This dataset has been annotated for aspect terms, opinion terms, polarities of the opinion terms towards targeted aspects, document-level opinion polarities and sentence separations. We develop a hierarchical taxonomy for aspect categorization, which covers all the areas of the teaching-learning process. We annotated both implicit and explicit aspects using this taxonomy. Annotation methodology, difficulties faced during the annotation, and the details about the aspect term categorization have been discussed in detail. This annotated corpus can be used for Aspect Extraction, Aspect Level Sentiment Analysis, and Document Level Sentiment Analysis. Also the baseline results for all three tasks are given in the paper.
Motivated by the sparsity of NLP resources for Eastern European languages, we present a broad index of existing Eastern European language resources (90+ datasets and 45+ models) published as a github repository open for updates from the community. Furthermore, to support the evaluation of commonsense reasoning tasks, we provide hand-crafted cross-lingual datasets for five different semantic tasks (namely news categorization, paraphrase detection, Natural Language Inference (NLI) task, tweet sentiment detection, and news sentiment detection) for some of the Eastern European languages. We perform several experiments with the existing multilingual models on these datasets to define the performance baselines and compare them to the existing results for other languages.
We present SuperGLUE benchmark adapted and translated into Slovene using a combination of human and machine translation. We describe the translation process and problems arising due to differences in morphology and grammar. We evaluate the translated datasets in several modes: monolingual, cross-lingual, and multilingual, taking into account differences between machine and human translated training sets. The results show that the monolingual Slovene SloBERTa model is superior to massively multilingual and trilingual BERT models, but these also show a good cross-lingual performance on certain tasks. The performance of Slovene models still lags behind the best English models.
In this paper we present two datasets for Tamasheq, a developing language mainly spoken in Mali and Niger. These two datasets were made available for the IWSLT 2022 low-resource speech translation track, and they consist of collections of radio recordings from daily broadcast news in Niger (Studio Kalangou) and Mali (Studio Tamani). We share (i) a massive amount of unlabeled audio data (671 hours) in five languages: French from Niger, Fulfulde, Hausa, Tamasheq and Zarma, and (ii) a smaller 17 hours parallel corpus of audio recordings in Tamasheq, with utterance-level translations in the French language. All this data is shared under the Creative Commons BY-NC-ND 3.0 license. We hope these resources will inspire the speech community to develop and benchmark models using the Tamasheq language.
This paper describes a method of semi-automatic word spotting in minority languages, from one and the same Aesop fable “The North Wind and the Sun” translated in Romance languages/dialects from Hexagonal (i.e. Metropolitan) France and languages from French Polynesia. The first task consisted of finding out how a dozen words such as “wind” and “sun” were translated in over 200 versions collected in the field — taking advantage of orthographic similarity, word position and context. Occurrences of the translations were then extracted from the phone-aligned recordings. The results were judged accurate in 96–97% of cases, both on the development corpus and a test set of unseen data. Corrected alignments were then mapped and basemaps were drawn to make various linguistic phenomena immediately visible. The paper exemplifies how regular expressions may be used for this purpose. The final result, which takes the form of an online speaking atlas (enriching the https://atlas.limsi.fr website), enables us to illustrate lexical, morphological or phonetic variation.
We present a Multilingual Open Text (MOT), a new multilingual corpus containing text in 44 languages, many of which have limited existing text resources for natural language processing. The first release of the corpus contains over 2.8 million news articles and an additional 1 million short snippets (photo captions, video descriptions, etc.) published between 2001–2022 and collected from Voice of America’s news websites. We describe our process for collecting, filtering, and processing the data. The source material is in the public domain, our collection is licensed using a creative commons license (CC BY 4.0), and all software used to create the corpus is released under the MIT License. The corpus will be regularly updated as additional documents are published.
Deploying recent natural language processing innovations to low-resource settings allows for state-of-the-art research findings and applications to be accessed across cultural and linguistic borders. One low-resource setting of increasing interest is code-switching, the phenomenon of combining, swapping, or alternating the use of two or more languages in continuous dialogue. In this paper, we introduce a large dataset (20k+ instances) to facilitate investigation of Tagalog-English code-switching, which has become a popular mode of discourse in Philippine culture. Tagalog is an Austronesian language and former official language of the Philippines spoken by over 23 million people worldwide, but it and Tagalog-English are under-represented in NLP research and practice. We describe our methods for data collection, as well as our labeling procedures. We analyze our resulting dataset, and finally conclude by providing results from a proof-of-concept regression task to establish dataset validity, achieving a strong performance benchmark (R2=0.797-0.909; RMSE=0.068-0.057).
This work presents a parallel corpus of Guarani-Spanish text aligned at sentence level. The corpus contains about 30,000 sentence pairs, and is structured as a collection of subsets from different sources, further split into training, development and test sets. A sample of sentences from the test set was manually annotated by native speakers in order to incorporate meta-linguistic annotations about the Guarani dialects present in the corpus and also the correctness of the alignment and translation. We also present some baseline MT experiments and analyze the results in terms of the subsets. We hope this corpus can be used as a benchmark for testing Guarani-Spanish MT systems, and aim to expand and improve the quality of the corpus in future iterations.
Although information on the Internet can be shared in many languages, the language presence on the World Wide Web is very disproportionate. The problem of multilingualism on the Web, in particular access, availability and quality of information in the world’s languages, has been the subject of UNESCO focus for several decades. Making European websites more multilingual is also one of the focal targets of the Connecting Europe Facility Automated Translation (CEF AT) digital service infrastructure. In order to monitor this goal, alongside other possible solutions, CEF AT needs a methodology and easy to use tool to assess the degree of multilingualism of a given website. In this paper we investigate methods and tools that automatically analyse the language diversity of the Web and propose indicators and methodology on how to measure the multilingualism of European websites. We also introduce a prototype tool based on open-source software that helps to assess multilingualism of the Web and can be independently run at set intervals. We also present initial results obtained with our tool that allows us to conclude that multilingualism on the Web is still a problem not only at the world level, but also at the European and regional level.
Different algorithms have been proposed to detect semantic shifts (changes in a word meaning over time) in a diachronic corpus. Yet, and somehow surprisingly, no reference corpus has been designed so far to evaluate them, leaving researchers to fallback to troublesome evaluation strategies. In this work, we introduce a methodology for the construction of a reference dataset for the evaluation of semantic shift detection, that is, a list of words where we know for sure whether they present a word meaning change over a period of interest. We leverage a state-of-the-art word-sense disambiguation model to associate a date of first appearance to all the senses of a word. Significant changes in sense distributions as well as clear stability are detected and the resulting words are inspected by experts using a dedicated interface before populating a reference dataset. As a proof of concept, we apply this methodology to a corpus of newspapers from Quebec covering the whole 20th century. We manually verified a subset of candidates, leading to QC-FR-Diac-V1.0, a corpus of 151 words allowing one to evaluate the identification of semantic shifts in French between 1910 and 1990.
We introduce the CRASS (counterfactual reasoning assessment) data set and benchmark utilizing questionized counterfactual conditionals as a novel and powerful tool to evaluate large language models. We present the data set design and benchmark. We test six state-of-the-art models against our benchmark. Our results show that it poses a valid challenge for these models and opens up considerable room for their improvement.
The scientific community is increasingly aware of the necessity to embrace pluralism and consistently represent major and minor social groups. Currently, there are no standard evaluation techniques for different types of biases. Accordingly, there is an urgent need to provide evaluation sets and protocols to measure existing biases in our automatic systems. Evaluating the biases should be an essential step towards mitigating them in the systems. This paper introduces WinoST, a new freely available challenge set for evaluating gender bias in speech translation. WinoST is the speech version of WinoMT, an MT challenge set, and both follow an evaluation protocol to measure gender accuracy. Using an S-Transformer end-to-end speech translation system, we report the gender bias evaluation on four language pairs, and we reveal the inaccuracies in translations generating gender-stereotyped translations.
Obtaining linguistic annotation from novice crowdworkers is far from trivial. A case in point is the annotation of discourse relations, which is a complicated task. Recent methods have obtained promising results by extracting relation labels from either discourse connectives (DCs) or question-answer (QA) pairs that participants provide. The current contribution studies the effect of worker selection and training on the agreement on implicit relation labels between workers and gold labels, for both the DC and the QA method. In Study 1, workers were not specifically selected or trained, and the results show that there is much room for improvement. Study 2 shows that a combination of selection and training does lead to improved results, but the method is cost- and time-intensive. Study 3 shows that a selection-only approach is a viable alternative; it results in annotations of comparable quality compared to annotations from trained participants. The results generalized over both the DC and QA method and therefore indicate that a selection-only approach could also be effective for other crowdsourced discourse annotation tasks.
Social media are heavily used by many users to share their mental health concerns and diagnoses. This trend has turned social media into a large-scale resource for researchers focused on detecting mental health conditions. Social media usage varies considerably across individuals. Thus, classification of patterns, including detecting signs of depression, must account for such variation. We address the disparity in classification effectiveness for users with little activity (e.g., new users). Our evaluation, performed on a large-scale dataset, shows considerable detection discrepancy based on user posting frequency. For instance, the F1 detection score of users with an above-median versus below-median number of posts is greater than double (0.803 vs 0.365) using a conventional CNN-based model; similar results were observed on lexical and transformer-based classifiers. To complement this evaluation, we propose a dynamic thresholding technique that adjusts the classifier’s sensitivity as a function of the number of posts a user has. This technique alone reduces the margin between users with many and few posts, on average, by 45% across all methods and increases overall performance, on average, by 33%. These findings emphasize the importance of evaluating and tuning natural language systems for potentially vulnerable populations.
Can NLP assist in building formal models for verifying complex systems? We study this challenge in the context of parsing Network File System (NFS) specifications. We define a semantic-dependency problem over SpecIR, a representation language we introduce to model sentences appearing in NFS specification documents (RFCs) as IF-THEN statements, and present an annotated dataset of 1,198 sentences. We develop and evaluate semantic-dependency parsing systems for this problem. Evaluations show that even when using a state-of-the-art language model, there is significant room for improvement, with the best models achieving an F1 score of only 60.5 and 33.3 in the named-entity-recognition and dependency-link-prediction sub-tasks, respectively. We also release additional unlabeled data and other domain-related texts. Experiments show that these additional resources increase the F1 measure when used for simple domain-adaption and transfer-learning-based approaches, suggesting fruitful directions for further research
NLP technologies such as text similarity assessment, question answering and text classification are increasingly being used to develop intelligent educational applications. The long-term goal of our work is an intelligent tutoring system for German secondary schools, which will support students in a school exercise that requires them to identify arguments in an argumentative source text. The present paper presents our work on a central subtask, viz. the automatic assessment of similarity between a pair of argumentative text snippets in German. In the designated use case, students write out key arguments from a given source text; the tutoring system then evaluates them against a target reference, assessing the similarity level between student work and the reference. We collect a dataset for our similarity assessment task through crowdsourcing as authentic German student data are scarce; we label the collected text pairs with similarity scores on a 5-point scale and run first experiments on the task. We see that a model based on BERT shows promising results, while we also discuss some challenges that we observe.
Studying and mitigating gender and other biases in natural language have become important areas of research from both algorithmic and data perspectives. This paper explores the idea of reducing gender bias in a language generation context by generating gender variants of sentences. Previous work in this field has either been rule-based or required large amounts of gender balanced training data. These approaches are however not scalable across multiple languages, as creating data or rules for each language is costly and time-consuming. This work explores a light-weight method to generate gender variants for a given text using pre-trained language models as the resource, without any task-specific labelled data. The approach is designed to work on multiple languages with minimal changes in the form of heuristics. To showcase that, we have tested it on a high-resourced language, namely Spanish, and a low-resourced language from a different family, namely Serbian. The approach proved to work very well on Spanish, and while the results were less positive for Serbian, it showed potential even for languages where pre-trained models are less effective.
Hate speech detection is a prominent and challenging task, since hate messages are often expressed in subtle ways and with characteristics that may vary depending on the author. Hence, many models suffer from the generalization problem. However, retrieving and monitoring hateful content on social media is a current necessity. In this paper, we propose an unsupervised approach using Graph Auto-Encoders (GAE), which allows us to avoid using labeled data when training the representation of the texts. Specifically, we represent texts as nodes of a graph, and use a transformer layer together with a convolutional layer to encode these nodes in a low-dimensional space. As a result, we obtain embeddings that can be decoded into a reconstruction of the original network. Our main idea is to learn a model with a set of texts without supervision, in order to generate embeddings for the nodes: nodes with the same label should be close in the embedding space, which, in turn, should allow us to distinguish among classes. We employ this strategy to detect hate speech in multi-domain and multilingual sets of texts, where our method shows competitive results on small datasets.
Question Answering, including Reading Comprehension, is one of the NLP research areas that has seen significant scientific breakthroughs over the past few years, thanks to the concomitant advances in Language Modeling. Most of these breakthroughs, however, are centered on the English language. In 2020, as a first strong initiative to bridge the gap to the French language, Illuin Technology introduced FQuAD1.1, a French Native Reading Comprehension dataset composed of 60,000+ questions and answers samples extracted from Wikipedia articles. Nonetheless, Question Answering models trained on this dataset have a major drawback: they are not able to predict when a given question has no answer in the paragraph of interest, therefore making unreliable predictions in various industrial use-cases. We introduce FQuAD2.0, which extends FQuAD with 17,000+ unanswerable questions, annotated adversarially, in order to be similar to answerable ones. This new dataset, comprising a total of almost 80,000 questions, makes it possible to train French Question Answering models with the ability of distinguishing unanswerable questions from answerable ones. We benchmark several models with this dataset: our best model, a fine-tuned CamemBERT-large, achieves a F1 score of 82.3% on this classification task, and a F1 score of 83% on the Reading Comprehension task.
The performance of hate speech detection models relies on the datasets on which the models are trained. Existing datasets are mostly prepared with a limited number of instances or hate domains that define hate topics. This hinders large-scale analysis and transfer learning with respect to hate domains. In this study, we construct large-scale tweet datasets for hate speech detection in English and a low-resource language, Turkish, consisting of human-labeled 100k tweets per each. Our datasets are designed to have equal number of tweets distributed over five domains. The experimental results supported by statistical tests show that Transformer-based language models outperform conventional bag-of-words and neural models by at least 5% in English and 10% in Turkish for large-scale hate speech detection. The performance is also scalable to different training sizes, such that 98% of performance in English, and 97% in Turkish, are recovered when 20% of training instances are used. We further examine the generalization ability of cross-domain transfer among hate domains. We show that 96% of the performance of a target domain in average is recovered by other domains for English, and 92% for Turkish. Gender and religion are more successful to generalize to other domains, while sports fail most.
Health behaviour change is a difficult and prolonged process that requires sustained motivation and determination. Conversa- tional agents have shown promise in supporting the change process in the past. One therapy approach that facilitates change and has been used as a framework for conversational agents is motivational interviewing. However, existing implementations of this therapy approach lack the deep understanding of user utterances that is essential to the spirit of motivational interviewing. To address this lack of understanding, we introduce the GLoHBCD, a German dataset of naturalistic language around health behaviour change. Data was sourced from a popular German weight loss forum and annotated using theoretically grounded motivational interviewing categories. We describe the process of dataset construction and present evaluation results. Initial experiments suggest a potential for broad applicability of the data and the resulting classifiers across different behaviour change domains. We make code to replicate the dataset and experiments available on Github.
Digital recorded written and spoken dialogues are becoming increasingly available as an effect of the technological advances such as online messenger services and the use of chatbots. Summaries are a natural way of presenting the important information gathered from dialogues. We present a unique data set that consists of Dutch spoken human-computer conversations, an annotation layer of turn labels, and conversational abstractive summaries of user answers. The data set is publicly available for research purposes.
Entity linking in dialogue is the task of mapping entity mentions in utterances to a target knowledge base. Prior work on entity linking has mainly focused on well-written articles such as Wikipedia, annotated newswire, or domain-specific datasets. We extend the study of entity linking to open domain dialogue by presenting the OpenEL corpus: an annotated multi-domain corpus for linking entities in natural conversation to Wikidata. Each dialogic utterance in 179 dialogues over 12 topics from the EDINA dataset has been annotated for entities realized by definite referring expressions as well as anaphoric forms such as he, she, it and they. This dataset supports training and evaluation of entity linking in open-domain dialogue, as well as analysis of the effect of using dialogue context and anaphora resolution in model training. It could also be used for fine-tuning a coreference resolution algorithm. To the best of our knowledge, this is the first substantial entity linking corpus publicly available for open-domain dialogue. We also establish baselines for this task using several existing entity linking systems. We found that the Transformer-based system Flair + BLINK has the best performance with a 0.65 F1 score. Our results show that dialogue context is extremely beneficial for entity linking in conversations, with Flair + Blink achieving an F1 of 0.61 without discourse context. These results also demonstrate the remaining performance gap between the baselines and human performance, highlighting the challenges of entity linking in open-domain dialogue, and suggesting many avenues for future research using OpenEL.
An idealized, though simplistic, view of the referring expression production and grounding process in (situated) dialogue assumes that a speaker must merely appropriately specify their expression so that the target referent may be successfully identified by the addressee. However, referring in conversation is a collaborative process that cannot be aptly characterized as an exchange of minimally-specified referring expressions. Concerns have been raised regarding assumptions made by prior work on visually-grounded dialogue that reveal an oversimplified view of conversation and the referential process. We address these concerns by introducing a collaborative image ranking task, a grounded agreement game we call “A Game Of Sorts”. In our game, players are tasked with reaching agreement on how to rank a set of images given some sorting criterion through a largely unrestricted, role-symmetric dialogue. By putting emphasis on the argumentation in this mixed-initiative interaction, we collect discussions that involve the collaborative referential process. We describe results of a small-scale data collection experiment with the proposed task. All discussed materials, which includes the collected data, the codebase, and a containerized version of the application, are publicly available.
This paper introduces CoRoSeOf, a large corpus of Romanian social media manually annotated for sexist and offensive language. We describe the annotation process of the corpus, provide initial analyses, and baseline classification results for sexism detection on this data set. The resulting corpus contains 39 245 tweets, annotated by multiple annotators (with an agreement rate of Fleiss’κ= 0.45), following the sexist label set of a recent study. The automatic sexism detection yields scores similar to some of the earlier studies (macro averaged F1 score of 83.07% on binary classification task). We release the corpus with a permissive license.
The use of misogynistic and sexist language has increased in recent years in social media, and is increasing in the Arabic world in reaction to reforms attempting to remove restrictions on women lives. However, there are few benchmarks for Arabic misogyny and sexism detection, and in those the annotations are in aggregated form even though misogyny and sexism judgments are found to be highly subjective. In this paper we introduce an Arabic misogyny and sexism dataset (ArMIS) characterized by providing annotations from annotators with different degree of religious beliefs, and provide evidence that such differences do result in disagreements. To the best of our knowledge, this is the first dataset to study in detail the effect of beliefs on misogyny and sexism annotation. We also discuss proof-of-concept experiments showing that a dataset in which disagreements have not been reconciled can be used to train state-of-the-art models for misogyny and sexism detection; and consider different ways in which such models could be evaluated.
Integrating the existing interruption and turn switch classification methods, we propose a new annotation schema to annotate different types of interruptions through timeliness, switch accomplishment and speech content level. The proposed method is able to distinguish smooth turn exchange, backchannel and interruption (including interruption types) and to annotate dyadic conversation. We annotated the French part of NoXi corpus with the proposed structure and use these annotations to study the probability distribution and duration of each turn switch type.
Despite the importance of understanding causality, corpora addressing causal relations are limited. There is a discrepancy between existing annotation guidelines of event causality and conventional causality corpora that focus more on linguistics. Many guidelines restrict themselves to include only explicit relations or clause-based arguments. Therefore, we propose an annotation schema for event causality that addresses these concerns. We annotated 3,559 event sentences from protest event news with labels on whether it contains causal relations or not. Our corpus is known as the Causal News Corpus (CNC). A neural network built upon a state-of-the-art pre-trained language model performed well with 81.20% F1 score on test set, and 83.46% in 5-folds cross-validation. CNC is transferable across two external corpora: CausalTimeBank (CTB) and Penn Discourse Treebank (PDTB). Leveraging each of these external datasets for training, we achieved up to approximately 64% F1 on the CNC test set without additional fine-tuning. CNC also served as an effective training and pre-training dataset for the two external corpora. Lastly, we demonstrate the difficulty of our task to the layman in a crowd-sourced annotation exercise. Our annotated corpus is publicly available, providing a valuable resource for causal text mining researchers.
This contribution describes the collection of a large and diverse corpus for speech recognition and similar tools using crowd-sourced donations. We have built a collection platform inspired by Mozilla Common Voice and specialized it to our needs. We discuss the importance of engaging the community and motivating it to contribute, in our case through competitions. Given the incentive and a platform to easily read in large amounts of utterances, we have observed four cases of speakers freely donating over 10 thousand utterances. We have also seen that women are keener to participate in these events throughout all age groups. Manually verifying a large corpus is a monumental task and we attempt to automatically verify parts of the data using tools like Marosijo and the Montreal Forced Aligner. The method proved helpful, especially for detecting invalid utterances and halving the work needed from crowd-sourced verification.
In recent years, machine learning for clinical decision support has gained more and more attention. In order to introduce such applications into clinical practice, a good performance might be essential, however, the aspect of trust should not be underestimated. For the treating physician using such a system and being (legally) responsible for the decision made, it is particularly important to understand the system’s recommendation. To provide insights into a model’s decision, various techniques from the field of explainability (XAI) have been proposed whose output is often enough not targeted to the domain experts that want to use the model. To close this gap, in this work, we explore how explanations could possibly look like in future. To this end, this work presents a dataset of textual explanations in context of decision support. Within a reader study, human physicians estimated the likelihood of possible negative patient outcomes in the near future and justified each decision with a few sentences. Using those sentences, we created a novel corpus, annotated with different semantic layers. Moreover, we provide an analysis of how those explanations are constructed, and how they change depending on physician, on the estimated risk and also in comparison to an automatic clinical decision support system with feature importance.
Disfluency detection is a critical task in real-time dialogue systems. However, despite its importance, it remains a relatively unexplored field, mainly due to the lack of appropriate datasets. At the same time, existing datasets suffer from various issues, including class imbalance issues, which can significantly affect the performance of the model on rare classes, as it is demonstrated in this paper. To this end, we propose LARD, a method for generating complex and realistic artificial disfluencies with little effort. The proposed method can handle three of the most common types of disfluencies: repetitions, replacements, and restarts. In addition, we release a new large-scale dataset with disfluencies that can be used on four different tasks: disfluency detection, classification, extraction, and correction. Experimental results on the LARD dataset demonstrate that the data produced by the proposed method can be effectively used for detecting and removing disfluencies, while also addressing limitations of existing datasets.
We describe a new freely available Chinese multi-party dialogue dataset for automatic extraction of dialogue-based character relationships. The data has been extracted from the original TV scripts of a Chinese sitcom called “I Love My Home” with complex family-based human daily spoken conversations in Chinese. First, we introduced human annotation scheme for both global Character relationship map and character reference relationship. And then we generated the dialogue-based character relationship triples. The corpus annotates relationships between 140 entities in total. We also carried out a data exploration experiment by deploying a BERT-based model to extract character relationships on the CRECIL corpus and another existing relation extraction corpus (DialogRE (CITATION)).The results demonstrate that extracting character relationships is more challenging in CRECIL than in DialogRE.
In recent years, the focus on developing natural language processing (NLP) tools for Arabic has shifted from Modern Standard Arabic to various Arabic dialects. Various corpora of various sizes and representing different genres, have been created for a number of Arabic dialects. As far as Gulf Arabic is concerned, Gumar Corpus (Khalifa et al., 2016) is the largest corpus, to date, that includes data representing the dialectal Arabic of the six Gulf Cooperation Council countries (Bahrain, Kuwait, Saudi Arabia, Qatar, United Arab Emirates, and Oman), particularly in the genre of “online forum novels”. In this paper, we present the Bahrain Corpus. Our objective is to create a specialized corpus of the Bahraini Arabic dialect, which includes written texts as well as transcripts of audio files, belonging to a different genre (folktales, comedy shows, plays, cooking shows, etc.). The corpus comprises 620K words, carefully curated. We provide automatic morphological annotations of the full corpus using state-of-the-art morphosyntactic disambiguation for Gulf Arabic. We validate the quality of the annotations on a 7.6K word sample. We plan to make the annotated sample as well as the full corpus publicly available to support researchers interested in Arabic NLP.
In this paper we present the initial construction of a Universal Dependencies treebank with morphological annotations of Ancient Hebrew containing portions of the Hebrew Scriptures (1579 sentences, 27K tokens) for use in comparative study with ancient translations and for analysis of the development of Hebrew syntax. We construct this treebank by applying a rule-based parser (300 rules) to an existing morphologically-annotated corpus with minimal constituency structure and manually verifying the output and present the results of this semi-automated annotation process and some of the annotation decisions made in the process of applying the UD guidelines to a new language.
This paper introduces FIGHT, a dataset containing 63,450 tweets, posted before and after the official declaration of Covid-19 as a pandemic by online users in Portugal. This resource aims at contributing to the analysis of online hate speech targeting the most representative minorities in Portugal, namely the African descent and the Roma communities, and the LGBTQI community, the most commonly reported target of hate speech in social media at the European context. We present the methods for collecting the data, and provide insightful statistics on the distribution of tweets included in FIGHT, considering both the temporal and spatial dimensions. We also analyze the availability over time of tweets targeting the above-mentioned communities, distinguishing public, private and deleted tweets. We believe this study will contribute to better understand the dynamics of online hate speech in Portugal, particularly in adverse contexts, such as a pandemic outbreak, allowing the development of more informed and accurate hate speech resources for Portuguese.
The Icelandic Gigaword Corpus was first published in 2018. Since then new versions have been published annually, containing new texts from additional sources as well as from previous sources. This paper describes the evolution of the corpus in its first four years. All versions are made available under permissive licenses and with each new version the texts are annotated with the latest and most accurate tools. We show how the corpus has grown almost 50% in size from the first version to the fourth and how it was restructured in order to better accommodate different meta-data for different subcorpora. Furthermore, other services have been set up to facilitate usage of the corpus for different use cases. These include a keyword-in-context concordance tool, an n-gram viewer, a word frequency database and pre-trained word embeddings.
New models for natural language understanding have recently made an unparalleled amount of progress, which has led some researchers to suggest that the models induce universal text representations. However, current benchmarks are predominantly targeting semantic phenomena; we make the case that pragmatics needs to take center stage in the evaluation of natural language understanding. We introduce PragmEval, a new benchmark for the evaluation of natural language understanding, that unites 11 pragmatics-focused evaluation datasets for English. PragmEval can be used as supplementary training data in a multi-task learning setup, and is publicly available, alongside the code for gathering and preprocessing the datasets. Using our evaluation suite, we show that natural language inference, a widely used pretraining task, does not result in genuinely universal representations, which presents a new challenge for multi-task learning.
Many socio-linguistic cues are used in conversational analysis, such as emotion, sentiment, and dialogue acts. One of the fundamental social cues is politeness, which linguistically possesses properties such as social manners useful in conversational analysis. This article presents findings of polite emotional dialogue act associations, where we can correlate the relationships between the socio-linguistic cues. We confirm our hypothesis that the utterances with the emotion classes Anger and Disgust are more likely to be impolite. At the same time, Happiness and Sadness are more likely to be polite. A less expectable phenomenon occurs with dialogue acts Inform and Commissive which contain more polite utterances than Question and Directive. Finally, we conclude on the future work of these findings to extend the learning of social behaviours using politeness.
Discourse marker inventories are important tools for the development of both discourse parsers and corpora with discourse annotations. In this paper we explore the potential of massively multilingual lexical knowledge graphs to induce multilingual discourse marker lexicons using concept propagation methods as previously developed in the context of translation inference across dictionaries. Given one or multiple source languages with discourse marker inventories that discourse relations as senses of potential discourse markers, as well as a large number of bilingual dictionaries that link them – directly or indirectly – with the target language, we specifically study to what extent discourse marker induction can benefit from the integration of information from different sources, the impact of sense granularity and what limiting factors may need to be considered. Our study uses discourse marker inventories from nine European languages normalized against the discourse relation inventory of the Penn Discourse Treebank (PDTB), as well as three collections of machine-readable dictionaries with different characteristics, so that the interplay of a large number of factors can be studied.
Topological Data Analysis (TDA) focuses on the inherent shape of (spatial) data. As such, it may provide useful methods to explore spatial representations of linguistic data (embeddings) which have become central in NLP. In this paper we aim to introduce TDA to researchers in language technology. We use TDA to represent document structure as so-called story trees. Story trees are hierarchical representations created from semantic vector representations of sentences via persistent homology. They can be used to identify and clearly visualize prominent components of a story line. We showcase their potential by using story trees to create extractive summaries for news stories.
The study of metaphors in media discourse is an increasingly researched topic as media are an important shaper of social reality and metaphors are an indicator of how we think about certain issues through references to other things. We present a neural transfer learning method for detecting metaphorical sentences in Slovene and evaluate its performance on a gold standard corpus of metaphors (classification accuracy of 0.725), as well as on a sample of a domain specific corpus of migrations (precision of 0.40 for extracting domain metaphors and 0.74 if evaluated only on a set of migration related sentences). Based on empirical results and findings of our analysis, we propose a novel metaphor annotation scheme containing linguistic level, conceptual level, and stance information. The new scheme can be used for future metaphor annotations of other socially relevant topics.
To date, there has been no resource for studying discourse coherence on real-world Danish texts. Discourse coherence has mostly been approached with the assumption that incoherent texts can be represented by coherent texts in which sentences have been shuffled. However, incoherent real-world texts rarely resemble that. We thus present DDisCo, a dataset including text from the Danish Wikipedia and Reddit annotated for discourse coherence. We choose to annotate real-world texts instead of relying on artificially incoherent text for training and testing models. Then, we evaluate the performance of several methods, including neural networks, on the dataset.
In argumentative discourse, persuasion is often achieved by refuting or attacking others’ arguments. Attacking an argument is not always straightforward and often consists of complex rhetorical moves in which arguers may agree with a logic of an argument while attacking another logic. Furthermore, an arguer may neither deny nor agree with any logics of an argument, instead ignore them and attack the main stance of the argument by providing new logics and presupposing that the new logics have more value or importance than the logics presented in the attacked argument. However, there are no studies in computational argumentation that capture such complex rhetorical moves in attacks or the presuppositions or value judgments in them. To address this gap, we introduce LPAttack, a novel annotation scheme that captures the common modes and complex rhetorical moves in attacks along with the implicit presuppositions and value judgments. Our annotation study shows moderate inter-annotator agreement, indicating that human annotation for the proposed scheme is feasible. We publicly release our annotated corpus and the annotation guidelines.
We present the BeSt corpus, which records cognitive state: who believes what (i.e., factuality), and who has what sentiment towards what. This corpus is inspired by similar source-and-target corpora, specifically MPQA and FactBank. The corpus comprises two genres, newswire and discussion forums, in three languages, Chinese (Mandarin), English, and Spanish. The corpus is distributed through the LDC.
MOTIF (MultimOdal ConTextualized Images For Language Learners) is a multimodal dataset that consists of 1125 comprehension texts retrieved from Wikipedia Simple Corpus. Allowing multimodal processing or enriching the context with multimodal information has proven imperative for many learning tasks, specifically for second language (L2) learning. In this respect, several traditional NLP approaches can assist L2 readers in text comprehension processes, such as simplifying text or giving dictionary descriptions for complex words. As nicely stated in the well-known proverb, sometimes “a picture is worth a thousand words” and an image can successfully complement the verbal message by enriching the representation, like in Pictionary books. This multimodal support can also assist on-the-fly text reading experience by providing a multimodal tool that chooses and displays the most relevant images for the difficult words, given the text context. This study mainly focuses on one of the key components to achieving this goal; collecting a multimodal dataset enriched with complex word annotation and validated image match.
Sign Languages (SLs) are the primary means of communication for at least half a million people in Europe alone. However, the development of SL recognition and translation tools is slowed down by a series of obstacles concerning resource scarcity and standardization issues in the available data. The former challenge relates to the volume of data available for machine learning as well as the time required to collect and process new data. The latter obstacle is linked to the variety of the data, i.e., annotation formats are not unified and vary amongst different resources. The available data formats are often not suitable for machine learning, obstructing the provision of automatic tools based on neural models. In the present paper, we give an overview of these challenges by comparing various SL corpora and SL machine learning datasets. Furthermore, we propose a framework to address the lack of standardization at format level, unify the available resources and facilitate SL research for different languages. Our framework takes ELAN files as inputs and returns textual and visual data ready to train SL recognition and translation models. We present a proof of concept, training neural translation models on the data produced by the proposed framework.
The automatic translation of sign language videos into transcribed texts is rarely approached in its whole, as it implies to finely model the grammatical mechanisms that govern these languages. The presented work is a first step towards the interpretation of French sign language (LSF) by specifically targeting iconicity and spatial referencing. This paper describes the LSF-SHELVES corpus as well as the original technology that was designed and implemented to collect it. Our goal is to use deep learning methods to circumvent the use of models in spatial referencing recognition. In order to obtain training material with sufficient variability, we designed a light-weight (and low-cost) capture protocol that enabled us to collect data from a large panel of LSF signers. This protocol involves the use of a portable device providing a 3D skeleton, and of a software developed specifically for this application to facilitate the post-processing of handshapes. The LSF-SHELVES includes simple and compound iconic and spatial dynamics, organized in 6 complexity levels, representing a total of 60 sequences signed by 15 LSF signers.
The Computational Linguistics Applications for Multimedia Services (CLAMS) platform provides access to computational content analysis tools for multimedia material. The version we present here is a robust update of an initial prototype implementation from 2019. The platform now sports a variety of image, video, audio and text processing tools that interact via a common multi-modal representation language named MMIF (Multi-Media Interchange Format). We describe the overall architecture, the MMIF format, some of the tools included in the platform, the process to set up and run a workflow, visualizations included in CLAMS, and evaluate aspects of the platform on data from the American Archive of Public Broadcasting, showing how CLAMS can add metadata to mass-digitized multimedia collections, metadata that are typically only available implicitly in now largely unsearchable digitized media in archives and libraries.
Given our society’s increased exposure to multimedia formats on social media platforms, efforts to understand how digital content impacts people’s emotions are burgeoning. As such, we introduce a U.S. gun violence news dataset that contains news headline and image pairings from 840 news articles with 15K high-quality, crowdsourced annotations on emotional responses to the news pairings. We created three experimental conditions for the annotation process: two with a single modality (headline or image only), and one multimodal (headline and image together). In contrast to prior works on affectively-annotated data, our dataset includes annotations on the dominant emotion experienced with the content, the intensity of the selected emotion and an open-ended, written component. By collecting annotations on different modalities of the same news content pairings, we explore the relationship between image and text influence on human emotional response. We offer initial analysis on our dataset, showing the nuanced affective differences that appear due to modality and individual factors such as political leaning and media consumption habits. Our dataset is made publicly available to facilitate future research in affective computing.
We present RoomReader, a corpus of multimodal, multiparty conversational interactions in which participants followed a collaborative student-tutor scenario designed to elicit spontaneous speech. The corpus was developed within the wider RoomReader Project to explore multimodal cues of conversational engagement and behavioural aspects of collaborative interaction in online environments. However, the corpus can be used to study a wide range of phenomena in online multimodal interaction. The publicly-shared corpus consists of over 8 hours of video and audio recordings from 118 participants in 30 gender-balanced sessions, in the “in-the-wild” online environment of Zoom. The recordings have been edited, synchronised, and fully transcribed. Student participants have been continuously annotated for engagement with a novel continuous scale. We provide questionnaires measuring engagement and group cohesion collected from the annotators, tutors and participants themselves. We also make a range of accompanying data available such as personality tests and behavioural assessments. The dataset and accompanying psychometrics present a rich resource enabling the exploration of a range of downstream tasks across diverse fields including linguistics and artificial intelligence. This could include the automatic detection of student engagement, analysis of group interaction and collaboration in online conversation, and the analysis of conversational behaviours in an online setting.
In this article, we present Quevedo, a software tool we have developed for the task of automatic processing of graphical languages. These are languages which use images to convey meaning, relying not only on the shape of symbols but also on their spatial arrangement in the page, and relative to each other. When presented in image form, these languages require specialized computational processing which is not the same as usually done either for natural language processing or for artificial vision. Quevedo enables this specialized processing, focusing on a data-based approach. As a command line application and library, it provides features for the collection and management of image datasets, and their machine learning recognition using neural networks and recognizer pipelines. This processing requires careful annotation of the source data, for which Quevedo offers an extensive and visual web-based annotation interface. In this article, we also briefly present a case study centered on the task of SignWriting recognition, the original motivation for writing the software. Quevedo is written in Python, and distributed freely under the Open Software License version 3.0.
We introduce the Merkel Podcast Corpus, an audio-visual-text corpus in German collected from 16 years of (almost) weekly Internet podcasts of former German chancellor Angela Merkel. To the best of our knowledge, this is the first single speaker corpus in the German language consisting of audio, visual and text modalities of comparable size and temporal extent. We describe the methods used with which we have collected and edited the data which involves downloading the videos, transcripts and other metadata, forced alignment, performing active speaker recognition and face detection to finally curate the single speaker dataset consisting of utterances spoken by Angela Merkel. The proposed pipeline is general and can be used to curate other datasets of similar nature, such as talk show contents. Through various statistical analyses and applications of the dataset in talking face generation and TTS, we show the utility of the dataset. We argue that it is a valuable contribution to the research community, in particular, due to its realistic and challenging material at the boundary between prepared and spontaneous speech.
This paper presents the methodology we used to crowdsource a data collection of a new large-scale signer independent dataset for Kazakh-Russian Sign Language (KRSL) created for Sign Language Processing. By involving the Deaf community throughout the research process, we firstly designed a research protocol and then performed an efficient crowdsourcing campaign that resulted in a new FluentSigners-50 dataset. The FluentSigners-50 dataset consists of 173 sentences performed by 50 KRSL signers for 43,250 video samples. Dataset contributors recorded videos in real-life settings on various backgrounds using various devices such as smartphones and web cameras. Therefore, each dataset contribution has a varying distance to the camera, camera angles and aspect ratio, video quality, and frame rates. Additionally, the proposed dataset contains a high degree of linguistic and inter-signer variability and thus is a better training set for recognizing a real-life signed speech. FluentSigners-50 is publicly available at https://krslproject.github.io/fluentsigners-50/
The Petit Larousse illustré is a French dictionary first published in 1905. Its division in two main parts on language and on history and geography corresponds to a major milestone in French lexicography as well as a repository of general knowledge from this period. Although the value of many entries from 1905 remains intact, some descriptions now have a dimension that is more historical than contemporary. They are nonetheless significant to analyze and understand cultural representations from this time. A comparison with more recent information or a verification of these entries would require a tedious manual work. In this paper, we describe a new lexical resource, where we connected all the dictionary entries of the history and geography part to current data sources. For this, we linked each of these entries to a wikidata identifier. Using the wikidata links, we can automate more easily the identification, comparison, and verification of historically-situated representations. We give a few examples on how to process wikidata identifiers and we carried out a small analysis of the entities described in the dictionary to outline possible applications. The resource, i.e. the annotation of 20,245 dictionary entries with wikidata links, is available from GitHub (https://github.com/pnugues/petit_larousse_1905/)
The paper presents current work on a German corpus annotated for metaphor. Metaphors denote entities or situations that are in some sense similar to the literal referent, e.g., when “Handschrift” ‘signature’ is used in the sense of ‘distinguishing mark’ or the suppression of hopes is introduced by the verb “verschütten” ‘bury’. The corpus is part of a project on register, hence, includes material from different registers that represent register variation along a number of important dimensions, but we believe that it is of interest to research on metaphor in general. The corpus extends previous annotation initiatives in that it not only annotates the metaphoric expressions themselves but also their respective relevant contexts that trigger a metaphorical interpretation of the expressions. For the corpus, we developed extended annotation guidelines, which specifically focus not only on the identification of these metaphoric contexts but also analyse in detail specific linguistic challenges for metaphor annotation that emerge due to the grammar of German.
We describe NorDiaChange: the first diachronic semantic change dataset for Norwegian. NorDiaChange comprises two novel subsets, covering about 80 Norwegian nouns manually annotated with graded semantic change over time. Both datasets follow the same annotation procedure and can be used interchangeably as train and test splits for each other. NorDiaChange covers the time periods related to pre- and post-war events, oil and gas discovery in Norway, and technological developments. The annotation was done using the DURel framework and two large historical Norwegian corpora. NorDiaChange is published in full under a permissive licence, complete with raw annotation data and inferred diachronic word usage graphs (DWUGs).
We explored transformer-based language models for ranking instances of Portuguese lexico-semantic relations. Weights were based on the likelihood of natural language sequences that transmitted the relation instances, and expectations were that they would be useful for filtering out noisier instances. However, after analysing the weights, no strong conclusions were taken. They are not correlated with redundancy, but are lower for instances with longer and more specific arguments, which may nevertheless be a consequence of their sensitivity to the frequency of such arguments. They did also not reveal to be useful when computing word similarity with network embeddings. Despite the negative results, we see the reported experiments and insights as another contribution for better understanding transformer language models like BERT and GPT, and we make the weighted instances publicly available for further research.
While contextual language models are now dominant in the field of Natural Language Processing, the representations they build at the token level are not always suitable for all uses. In this article, we propose a new method for building word or type-level embeddings from contextual models. This method combines the generalization and the aggregation of token representations. We evaluate it for a large set of English nouns from the perspective of the building of distributional thesauri for extracting semantic similarity relations. Moreover, we analyze the differences between static embeddings and type-level embeddings according to features such as the frequency of words or the type of semantic relations these embeddings account for, showing that the properties of these two types of embeddings can be complementary and exploited for further improving distributional thesauri.
Many applications crucially rely on the availability of high-quality word vectors. To learn such representations, several strategies based on language models have been proposed in recent years. While effective, these methods typically rely on a large number of contextualised vectors for each word, which makes them impractical. In this paper, we investigate whether similar results can be obtained when only a few contextualised representations of each word can be used. To this end, we analyse a range of strategies for selecting the most informative sentences. Our results show that with a careful selection strategy, high-quality word vectors can be learned from as few as 5 to 10 sentences.
We provide a novel dataset – DiaWUG – with judgements on diatopic lexical semantic variation for six Spanish variants in Europe and Latin America. In contrast to most previous meaning-based resources and studies on semantic diatopic variation, we collect annotations on semantic relatedness for Spanish target words in their contexts from both a semasiological perspective (i.e., exploring the meanings of a word given its form, thus including polysemy) and an onomasiological perspective (i.e., exploring identical meanings of words with different forms, thus including synonymy). In addition, our novel dataset exploits and extends the existing framework DURel for annotating word senses in context (Erk et al., 2013; Schlechtweg et al., 2018) and the framework-embedded Word Usage Graphs (WUGs) – which up to now have mainly be used for semasiological tasks and resources – in order to distinguish, visualize and interpret lexical semantic variation of contextualized words in Spanish from these two perspectives, i.e., semasiological and onomasiological language variation.
Adpositions and case markers contain a high degree of polysemy and participate in unique semantic role configurations. We present a novel application of the SNACS supersense hierarchy to Finnish and Latin data by manually annotating adposition and case marker tokens in Finnish and Latin translations of Chapters IV-V of Le Petit Prince (The Little Prince). We evaluate the computational validity of the semantic role annotation categories by grouping raw, contextualized Multilingual BERT embeddings using k-means clustering.
Target Sense Verification (TSV) describes the binary disambiguation task of deciding whether the intended sense of a target word in a context corresponds to a given target sense. In this paper, we introduce WiC-TSV-de, a multi-domain dataset for German Target Sense Verification. While the training and development sets consist of domain-independent instances only, the test set contains domain-bound subsets, originating from four different domains, being Gastronomy, Medicine, Hunting, and Zoology. The domain-bound subsets incorporate adversarial examples such as in-domain ambiguous target senses and context-mixing (i.e., using the target sense in an out-of-domain context) which contribute to the challenging nature of the presented dataset. WiC-TSV-de allows for the development of sense-inventory-independent disambiguation models that can generalise their knowledge for different domain settings. By combining it with the original English WiC-TSV benchmark, we performed monolingual and cross-lingual analysis, where the evaluated baseline models were not able to solve the dataset to a satisfying degree, leaving a big gap to human performance.
BERT models used in specialized domains all seem to be the result of a simple strategy: initializing with the original BERT and then resuming pre-training on a specialized corpus. This method yields rather good performance (e.g. BioBERT (Lee et al., 2020), SciBERT (Beltagy et al., 2019), BlueBERT (Peng et al., 2019)). However, it seems reasonable to think that training directly on a specialized corpus, using a specialized vocabulary, could result in more tailored embeddings and thus help performance. To test this hypothesis, we train BERT models from scratch using many configurations involving general and medical corpora. Based on evaluations using four different tasks, we find that the initial corpus only has a weak influence on the performance of BERT models when these are further pre-trained on a medical corpus.
In this paper, we present the Universal Semantic Annotator (USeA), which offers the first unified API for high-quality automatic annotations of texts in 100 languages through state-of-the-art systems for Word Sense Disambiguation, Semantic Role Labeling and Semantic Parsing. Together, such annotations can be used to provide users with rich and diverse semantic information, help second-language learners, and allow researchers to integrate explicit semantic knowledge into downstream tasks and real-world applications.
DBLP is the largest open-access repository of scientific articles on computer science and provides metadata associated with publications, authors, and venues. We retrieved more than 6 million publications from DBLP and extracted pertinent metadata (e.g., abstracts, author affiliations, citations) from the publication texts to create the DBLP Discovery Dataset (D3). D3 can be used to identify trends in research activity, productivity, focus, bias, accessibility, and impact of computer science research. We present an initial analysis focused on the volume of computer science research (e.g., number of papers, authors, research activity), trends in topics of interest, and citation patterns. Our findings show that computer science is a growing research field (15% annually), with an active and collaborative researcher community. While papers in recent years present more bibliographical entries in comparison to previous decades, the average number of citations has been declining. Investigating papers’ abstracts reveals that recent topic trends are clearly reflected in D3. Finally, we list further applications of D3 and pose supplemental research questions. The D3 dataset, our findings, and source code are publicly available for research purposes.
This paper presents SciPar, a new collection of parallel corpora created from openly available metadata of bachelor theses, master theses and doctoral dissertations hosted in institutional repositories, digital libraries of universities and national archives. We describe first how we harvested and processed metadata from 86, mainly European, repositories to extract bilingual titles and abstracts, and then how we mined high quality sentence pairs in a wide range of scientific areas and sub-disciplines. In total, the resource includes 9.17 million segment alignments in 31 language pairs and is publicly available via the ELRC-SHARE repository. The bilingual corpora in this collection could prove valuable in various applications, such as cross-lingual plagiarism detection or adapting Machine Translation systems for the translation of scientific texts and academic writing in general, especially for language pairs which include English.
Euphemisms have not received much attention in natural language processing, despite being an important element of polite and figurative language. Euphemisms prove to be a difficult topic, not only because they are subject to language change, but also because humans may not agree on what is a euphemism and what is not. Nonetheless, the first step to tackling the issue is to collect and analyze examples of euphemisms. We present a corpus of potentially euphemistic terms (PETs) along with example texts from the GloWbE corpus. Additionally, we present a subcorpus of texts where these PETs are not being used euphemistically, which may be useful for future applications. We also discuss the results of multiple analyses run on the corpus. Firstly, we find that sentiment analysis on the euphemistic texts supports that PETs generally decrease negative and offensive sentiment. Secondly, we observe cases of disagreement in an annotation task, where humans are asked to label PETs as euphemistic or not in a subset of our corpus text examples. We attribute the disagreement to a variety of potential reasons, including if the PET was a commonly accepted term (CAT).
We present the Camel Treebank (CAMELTB), a 188K word open-source dependency treebank of Modern Standard and Classical Arabic. CAMELTB 1.0 includes 13 sub-corpora comprising selections of texts from pre-Islamic poetry to social media online commentaries, and covering a range of genres from religious and philosophical texts to news, novels, and student essays. The texts are all publicly available (out of copyright, creative commons, or under open licenses). The texts were morphologically tokenized and syntactically parsed automatically, and then manually corrected by a team of trained annotators. The annotations follow the guidelines of the Columbia Arabic Treebank (CATiB) dependency representation. We discuss our annotation process and guideline extensions, and we present some initial observations on lexical and syntactic differences among the annotated sub-corpora. This corpus will be publicly available to support and encourage research on Arabic NLP in general and on new, previously unexplored genres that are of interest to a wider spectrum of researchers, from historical linguistics and digital humanities to computer-assisted language pedagogy.
Mental health remains a significant challenge of public health worldwide. With increasing popularity of online platforms, many use the platforms to share their mental health conditions, express their feelings, and seek help from the community and counselors. Some of these platforms, such as Reachout, are dedicated forums where the users register to seek help. Others such as Reddit provide subreddits where the users publicly but anonymously post their mental health distress. Although posts are of varying length, it is beneficial to provide a short, but informative summary for fast processing by the counselors. To facilitate research in summarization of mental health online posts, we introduce Mental Health Summarization dataset, MentSum, containing over 24k carefully selected user posts from Reddit, along with their short user-written summary (called TLDR) in English from 43 mental health subreddits. This domain-specific dataset could be of interest not only for generating short summaries on Reddit, but also for generating summaries of posts on the dedicated mental health forums such as Reachout. We further evaluate both extractive and abstractive state-of-the-art summarization baselines in terms of Rouge scores, and finally conduct an in-depth human evaluation study of both user-written and system-generated summaries, highlighting challenges in this research.
Traditionally, Text Simplification is treated as a monolingual translation task where sentences between source texts and their simplified counterparts are aligned for training. However, especially for longer input documents, summarizing the text (or dropping less relevant content altogether) plays an important role in the simplification process, which is currently not reflected in existing datasets. Simultaneously, resources for non-English languages are scarce in general and prohibitive for training new solutions. To tackle this problem, we pose core requirements for a system that can jointly summarize and simplify long source documents. We further describe the creation of a new dataset for joint Text Simplification and Summarization based on German Wikipedia and the German children’s encyclopedia “Klexikon”, consisting of almost 2,900 documents. We release a document-aligned version that particularly highlights the summarization aspect, and provide statistical evidence that this resource is well suited to simplification as well. Code and data are available on Github: https://github.com/dennlinger/klexikon
The distribution of fake news is not a new but a rapidly growing problem. The shift to news consumption via social media has been one of the drivers for the spread of misleading and deliberately wrong information, as in addition to its ease of use there is rarely any veracity monitoring. Due to the harmful effects of such fake news on society, the detection of these has become increasingly important. We present an approach to the problem that combines the power of transformer-based language models while simultaneously addressing one of their inherent problems. Our framework, CMTR-BERT, combines multiple text representations, with the goal of circumventing sequential limits and related loss of information the underlying transformer architecture typically suffers from. Additionally, it enables the incorporation of contextual information. Extensive experiments on two very different, publicly available datasets demonstrates that our approach is able to set new state-of-the-art performance benchmarks. Apart from the benefit of using automatic text summarization techniques we also find that the incorporation of contextual information contributes to performance gains.
The CLARIN Concept Registry (CCR) is the common semantic ground for most CMDI-based profiles to describe language-related resources in the CLARIN universe. While the CCR supports semantic interoperability within this universe, it does not extend beyond it. The flexibility of CMDI, however, allows users to use other term or concept registries when defining their metadata components. In this paper, we describe our use of schema.org, a light ontology used by many parties across disciplines.
The QUEST (QUality ESTablished) project aims at ensuring the reusability of audio-visual datasets (Wamprechtshammer et al., 2022) by devising quality criteria and curating processes. RefCo (Reference Corpora) is an initiative within QUEST in collaboration with DoReCo (Documentation Reference Corpus, Paschen et al. (2020)) focusing on language documentation projects. Previously, Aznar and Seifart (2020) introduced a set of quality criteria dedicated to documenting fieldwork corpora. Based on these criteria, we establish a semi-automatic review process for existing and work-in-progress corpora, in particular for language documentation. The goal is to improve the quality of a corpus by increasing its reusability. A central part of this process is a template for machine-readable corpus documentation and automatic data verification based on this documentation. In addition to the documentation and automatic verification, the process involves a human review and potentially results in a RefCo certification of the corpus. For each of these steps, we provide guidelines and manuals. We describe the evaluation process in detail, highlight the current limits for automatic evaluation and how the manual review is organized accordingly.
Despite the recent findings on the conceptual and linguistic organization of personification, we have relatively little knowledge about its lexical patterns and grammatical templates. It is especially true in the case of Hungarian which has remained an understudied language regarding the constructions of figurative meaning generation. The present paper aims to provide a corpus-driven approach to personification analysis in the framework of cognitive linguistics. This approach is based on the building of a semi-automatically processed research corpus (the PerSE corpus) in which personifying linguistic structures are annotated manually. The present test version of the corpus consists of online car reviews written in Hungarian (10468 words altogether): the texts were tokenized, lemmatized, morphologically analyzed, syntactically parsed, and PoS-tagged with the e-magyar NLP tool. For the identification of personifications, the adaptation of the MIPVU protocol was used and combined with additional analysis of semantic relations within personifying multi-word expressions. The paper demonstrates the structure of the corpus as well as the levels of the annotation. Furthermore, it gives an overview of possible data types emerging from the analysis: lexical pattern, grammatical characteristics, and the construction-like behavior of personifications in Hungarian.
Discourse markers carry information about the discourse structure and organization, and also signal local dependencies or epistemological stance of speaker. They provide instructions on how to interpret the discourse, and their study is paramount to understand the mechanism underlying discourse organization. This paper presents a new language resource, an ISO-based annotated multilingual parallel corpus for discourse markers. The corpus comprises nine languages, Bulgarian, Lithuanian, German, European Portuguese, Hebrew, Romanian, Polish, and Macedonian, with English as a pivot language. In order to represent the meaning of the discourse markers, we propose an annotation scheme of discourse relations from ISO 24617-8 with a plug-in to ISO 24617-2 for communicative functions. We describe an experiment in which we applied the annotation scheme to assess its validity. The results reveal that, although some extensions are required to cover all the multilingual data, it provides a proper representation of discourse markers value. Additionally, we report some relevant contrastive phenomena concerning discourse markers interpretation and role in discourse. This first step will allow us to develop deep learning methods to identify and extract discourse relations and communicative functions, and to represent that information as Linguistic Linked Open Data (LLOD).
Speech is considered as a multi-modal process where hearing and vision are two fundamentals pillars. In fact, several studies have demonstrated that the robustness of Automatic Speech Recognition systems can be improved when audio and visual cues are combined to represent the nature of speech. In addition, Visual Speech Recognition, an open research problem whose purpose is to interpret speech by reading the lips of the speaker, has been a focus of interest in the last decades. Nevertheless, in order to estimate these systems in the currently Deep Learning era, large-scale databases are required. On the other hand, while most of these databases are dedicated to English, other languages lack sufficient resources. Thus, this paper presents a semi-automatically annotated audiovisual database to deal with unconstrained natural Spanish, providing 13 hours of data extracted from Spanish television. Furthermore, baseline results for both speaker-dependent and speaker-independent scenarios are reported using Hidden Markov Models, a traditional paradigm that has been widely used in the field of Speech Technologies.
Video-and-Language learning, such as video question answering or video captioning, is the next challenge in the deep learning society, as it pursues the way how human intelligence perceives everyday life. These tasks require the ability of multi-modal reasoning which is to handle both visual information and text information simultaneously across time. In this point of view, a cross-modality attention module that fuses video representation and text representation takes a critical role in most recent approaches. However, existing Video-and-Language models merely compute the attention weights without considering the different characteristics of video modality and text modality. Such na ̈ıve attention module hinders the current models to fully enjoy the strength of cross-modality. In this paper, we propose a novel Modality Alignment method that benefits the cross-modality attention module by guiding it to easily amalgamate multiple modalities. Specifically, we exploit Centered Kernel Alignment (CKA) which was originally proposed to measure the similarity between two deep representations. Our method directly optimizes CKA to make an alignment between video and text embedding representations, hence it aids the cross-modality attention module to combine information over different modalities. Experiments on real-world Video QA tasks demonstrate that our method outperforms conventional multi-modal methods significantly with +3.57% accuracy increment compared to the baseline in a popular benchmark dataset. Additionally, in a synthetic data environment, we show that learning the alignment with our method boosts the performance of the cross-modality attention.
This paper investigates the correlation between mutual gaze and linguistic repetition, a form of alignment, which we take as evidence of mutual understanding. We focus on a multimodal corpus made of three-party conversations and explore the question of whether mutual gaze events correspond to moments of repetition or non-repetition. Our results, although mainly significant on word unigrams and bigrams, suggest positive correlations between the presence of mutual gaze and the repetitions of tokens, lemmas, or parts-of-speech, but negative correlations when it comes to paired levels of representation (tokens or lemmas associated with their part-of-speech). No compelling correlation is found with duration of mutual gaze. Results are strongest when ignoring punctuation as representations of pauses, intonation, etc. in counting aligned tokens.
In natural language settings, many interactions include more than two speakers, and real-life interpretation is based on all types of information available in all modalities. This constitutes a challenge for corpus-based analyses because the information in the audio and visual channels must be included in the coding. The goal of the DINLANG project is to tackle that challenge and analyze spontaneous interactions in family dinner settings (two adults and two to three children). The families use either French, or LSF (French sign language). Our aim is to compare how participants share language across the range of modalities found in vocal and visual languaging in coordination with dining. In order to pinpoint similarities and differences, we had to find a common coding tool for all situations (variations from one family to another) and modalities. Our coding procedure incorporates the use of the ELAN software. We created a template organized around participants, situations, and modalities, rather than around language forms. Spoken language transcription can be integrated, when it exists, but it is not mandatory. Data that has been created with another software can be injected in ELAN files if it is linked using time stamps. Analyses performed with the coded files rely on ELAN’s structured search functionalities, which allow to achieve fine-grained temporal analyses and which can be completed by using spreadsheets or R language.
Words of any language are to some extent related thought the ways they are formed. For instance, the verb ‘exempl-ify’ and the noun ‘example-s’ are both based on the word ‘example’, but the verb is derived from it, while the noun is inflected. In Natural Language Processing of Russian, the inflection is satisfactorily processed; however, there are only a few machine-trackable resources that capture derivations even though Russian has both of these morphological processes very rich. Therefore, we devote this paper to improving one of the methods of constructing such resources and to the application of the method to a Russian lexicon, which results in the creation of the largest lexical resource of Russian derivational relations. The resulting database dubbed DeriNet.RU includes more than 300 thousand lexemes connected with more than 164 thousand binary derivational relations. To create such data, we combined the existing machine-learning methods that we improved to manage this goal. The whole approach is evaluated on our newly created data set of manual, parallel annotation. The resulting DeriNet.RU is freely available under an open license agreement.
This paper describes a method to enrich lexical resources with content relating to linguistic diversity, based on knowledge from the field of lexical typology. We capture the phenomenon of diversity through the notion of lexical gap and use a systematic method to infer gaps semi-automatically on a large scale, which we demonstrate on the kinship domain. The resulting free diversity-aware terminological resource consists of 198 concepts, 1,911 words, and 37,370 gaps in 699 languages. We see great potential in the use of resources such as ours for the improvement of a variety of cross-lingual NLP tasks, which we illustrate through an application in the evaluation of machine translation systems.
In this paper we describe our current work on creating a WordNet for Latvian based on the principles of the Princeton WordNet. The chosen methodology for word sense definition and sense linking is based on corpus evidence and the existing Tezaurs.lv online dictionary, ensuring a foundation that fits the Latvian language usage and existing linguistic tradition. We cover a wide set of semantic relations, including gradation sets. Currently the dataset consists of 6432 words linked in 5528 synsets, out of which 2717 synsets are considered fully completed as they have all the outgoing semantic links annotated, annotated with corpus examples for each sense and links to the English Princeton WordNet.
This paper presents a simple but effective method to build sentiment lexicons for the three Mainland Scandinavian languages: Danish, Norwegian and Swedish. This method benefits from the English Sentiwordnet and a thesaurus in one of the target languages. Sentiment information from the English resource is mapped to the target languages by using machine translation and similarity measures based on sentence embeddings. A number of experiments with Scandinavian languages are performed in order to determine the best working sentence embedding algorithm for this task. A careful extrinsic evaluation on several datasets yields state-of-the-art results using a simple rule-based sentiment analysis algorithm. The resources are made freely available under an MIT License.
This paper describes how a newly published Danish sentiment lexicon with a high lexical coverage was compiled by use of lexicographic methods and based on the links between groups of words listed in semantic order in a thesaurus and the corresponding word sense descriptions in a comprehensive monolingual dictionary. The overall idea was to identify negative and positive sections in a thesaurus, extract the words from these sections and combine them with the dictionary information via the links. The annotation task of the dataset included several steps, and was based on the comparison of synonyms and near synonyms within a semantic field. In the cases where one of the words were included in the smaller Danish sentiment lexicon AFINN, its value there was used as inspiration and expanded to the synonyms when appropriate. In order to obtain a more practical lexicon with overall polarity values at lemma level, all the senses of the lemma were afterwards compared, taking into consideration dictionary information such as usage, style and frequency. The final lexicon contains 13,859 Danish polarity lemmas and includes morphological information. It is freely available at https://github.com/dsldk/danish-sentiment-lexicon (licence CC-BY-SA 4.0 International).
We introduce the IndoUKC, a new multilingual lexical database comprised of eighteen Indian languages, with a focus on formally capturing words and word meanings specific to Indian languages and cultures. The IndoUKC reuses content from the existing IndoWordNet resource while providing a new model for the cross-lingual mapping of lexical meanings that allows for a richer, diversity-aware representation. Accordingly, beyond a thorough syntactic and semantic cleaning, the IndoWordNet lexical content has been thoroughly remodeled in order to allow a more precise expression of language-specific meaning. The resulting database is made available both for browsing through a graphical web interface and for download through the LiveLanguage data catalogue.
While pre-trained language models play a vital role in modern language processing tasks, but not every language can benefit from them. Most existing research on pre-trained language models focuses primarily on widely-used languages such as English, Chinese, and Indo-European languages. Additionally, such schemes usually require extensive computational resources alongside a large amount of data, which is infeasible for less-widely used languages. We aim to address this research niche by building a language model that understands the linguistic phenomena in the target language which can be trained with low-resources. In this paper, we discuss Korean language modeling, specifically methods for language representation and pre-training methods. With our Korean-specific language representation, we are able to build more powerful language models for Korean understanding, even with fewer resources. The paper proposes chunk-wise reconstruction of the Korean language based on a widely used transformer architecture and bidirectional language representation. We also introduce morphological features such as Part-of-Speech (PoS) into the language understanding by leveraging such information during the pre-training. Our experiment results prove that the proposed methods improve the model performance of the investigated Korean language understanding tasks.
Whole-person functional limitations in the areas of mobility, self-care and domestic life affect a majority of individuals with disabilities. Detecting, recording and monitoring such limitations would benefit those individuals, as well as research on whole-person functioning and general public health. Dictionaries of terms related to whole-person function would enable automated identification and extraction of relevant information. However, no such terminologies currently exist, due in part to a lack of standardized coding and their availability mainly in free text clinical notes. In this paper, we introduce terminologies of whole-person function in the domains of mobility, self-care and domestic life, built and evaluated using a small set of manually annotated clinical notes, which provided a seedset that was expanded using a mix of lexical and deep learning approaches.
The paper gives an account of an infrastructure that will be integrated into a platform aimed at providing a multi-faceted experience to visitors of Northern Greece using mythology as a starting point. This infrastructure comprises a multi-lingual and multi-modal corpus (i.e., a corpus of textual data supplemented with images, and video) that belongs to the humanities domain along with a dedicated database (content management system) with advanced indexing, linking and search functionalities. We will present the corpus itself focusing on the content, the methodology adopted for its development, and the steps taken towards rendering it accessible via the database in a way that also facilitates useful visualizations. In this context, we tried to address three main challenges: (a) to add a novel annotation layer, namely geotagging, (b) to ensure the long-term maintenance of and accessibility to the highly heterogeneous primary data – even after the life cycle of the current project – by adopting a metadata schema that is compatible to existing standards; and (c) to render the corpus a useful resource to scholarly research in the digital humanities by adding a minimum set of linguistic annotations.
The present data formats proposed by authentic organizations are based on a so-called standoff-style data format in XML, which represents a semantic data model through an instance structure and a link structure. However, this type of data formats intended to enhance the power of representation of an XML format injures the mobility of data because an abstract data structure denoted by multiple link paths is hard to be converted into other data structures. This difficulty causes a problem in the reuse of data to convert into other data formats especially in a personal data management environment. In this paper, in order to compensate for the drawback, we propose a new concept of transforming a link structure to an instance structure on a new marked-up scheme. This approach to language data brings a new architecture of language data management to realize a personal data management environment in daily and long-life use.
Political authorities in democratic countries regularly consult the public in order to allow citizens to voice their ideas and concerns on specific issues. When trying to evaluate the (often large number of) contributions by the public in order to inform decision-making, authorities regularly face challenges due to restricted resources. We identify several tasks whose automated support can help in the evaluation of public participation. These are i) the recognition of arguments, more precisely premises and their conclusions, ii) the assessment of the concreteness of arguments, iii) the detection of textual descriptions of locations in order to assign citizens’ ideas to a spatial location, and iv) the thematic categorization of contributions. To enable future research efforts to develop techniques addressing these four tasks, we introduce the CIMT PartEval Corpus, a new publicly-available German-language corpus that includes several thousand citizen contributions from six mobility-related planning processes in five German municipalities. The corpus provides annotations for each of these tasks which have not been available in German for the domain of public participation before either at all or in this scope and variety.
Typological databases can contain a wealth of information beyond the collection of linguistic properties across languages. This paper shows how information often overlooked in typological databases can inform the research community about the state of description of the world’s languages. We illustrate this using Grambank, a morphosyntactic typological database covering 2,467 language varieties and based on 3,951 grammatical descriptions. We classify and quantify the comments that accompany coded values in Grambank. We then aggregate these comments and the coded values to derive a level of description for 17 grammatical domains that Grambank covers (negation, adnominal modification, participant marking, tense, aspect, etc.). We show that the description level of grammatical domains varies across space and time. Information about gaps and uncertainties in the descriptive knowledge of grammatical domains within and across languages is essential for a correct analysis of data in typological databases and for the study of grammatical diversity more generally. When collected in a database, such information feeds into disciplines that focus on primary data collection, such as grammaticography and language documentation.
This paper showcases the utility and timeliness of the Hong Kong Protest News Dataset, a highly curated collection of news articles from diverse news sources, to investigate longitudinal and synchronic news characterisations of protests in Hong Kong between 1998 and 2020. The properties of the dataset enable us to apply natural language processing to its 4522 articles and thereby study patterns of journalistic practice across newspapers. This paper sheds light on whether depth and/or manner of reporting changed over time, and if so, in what ways, or in response to what. In its focus and methodology, this paper helps bridge the gap between “validity-focused methodological debates” and the use of computational methods of analysis in the social sciences.
This paper is based on a collection of 16th century letters from and to the Zurich reformer Heinrich Bullinger. Around 12,000 letters of this exchange have been preserved, out of which 3100 have been professionally edited, and another 5500 are available as provisional transcriptions. We have investigated code-switching in these 8600 letters, first on the sentence-level and then on the word-level. In this paper we give an overview of the corpus and its language mix (mostly Early New High German and Latin, but also French, Greek, Italian and Hebrew). We report on our experiences with a popular language identifier and present our results when training an alternative identifier on a very small training corpus of only 150 sentences per language. We use the automatically labeled sentences in order to bootstrap a word-based language classifier which works with high accuracy. Our research around the corpus building and annotation involves automatic handwritten text recognition, text normalisation for ENH German, and machine translation from medieval Latin into modern German.
This paper presents an analysis of annotation using an automatic pre-annotation for a mid-level annotation complexity task - dependency syntax annotation. It compares the annotation efforts made by annotators using a pre-annotated version (with a high-accuracy parser) and those made by fully manual annotation. The aim of the experiment is to judge the final annotation quality when pre-annotation is used. In addition, it evaluates the effect of automatic linguistically-based (rule-formulated) checks and another annotation on the same data available to the annotators, and their influence on annotation quality and efficiency. The experiment confirmed that the pre-annotation is an efficient tool for faster manual syntactic annotation which increases the consistency of the resulting annotation without reducing its quality.
TimeML is an annotation scheme for capturing temporal information in text. The developers of TimeML built the TimeBank corpus to both validate the scheme and provide a rich dataset of events, temporal expressions, and temporal relationships for training and testing temporal analysis systems. In our own work we have been developing methods aimed at TimeML graphs for detecting (and eventually automatically correcting) temporal inconsistencies, extracting timelines, and assessing temporal indeterminacy. In the course of this investigation we identified numerous previously unrecognized issues in the TimeBank corpus, including multiple violations of TimeML annotation guide rules, incorrectly disconnected temporal graphs, as well as inconsistent, redundant, missing, or otherwise incorrect annotations. We describe our methods for detecting and correcting these problems, which include: (a) automatic guideline checking (109 violations); (b) automatic inconsistency checking (65 inconsistent files); (c) automatic disconnectivity checking (625 incorrect breakpoints); and (d) manual comparison with the output of state-of-the-art automatic annotators to identify missing annotations (317 events, 52 temporal expressions). We provide our code as well as a set of patch files that can be applied to the TimeBank corpus to produce a corrected version for use by other researchers in the field.
In this paper, we present an evaluation of sentence representation models on the paraphrase detection task. The evaluation is designed to simulate a real-world problem of plagiarism and is based on one of the most important cases of forgery in modern history: the so-called “Protocols of the Elders of Zion”. The sentence pairs for the evaluation are taken from the infamous forged text “Protocols of the Elders of Zion” (Protocols) by unknown authors; and by “Dialogue in Hell between Machiavelli and Montesquieu” by Maurice Joly. Scholars have demonstrated that the first text plagiarizes from the second, indicating all the forged parts on qualitative grounds. Following this evidence, we organized the rephrased texts and asked native speakers to quantify the level of similarity between each pair. We used this material to evaluate sentence representation models in two languages: English and French, and on three tasks: similarity correlation, paraphrase identification, and paraphrase retrieval. Our evaluation aims at encouraging the development of benchmarks based on real-world problems, as a means to prevent problems connected to AI hypes, and to use NLP technologies for social good. Through our evaluation, we are able to confirm that the infamous Protocols are actually a plagiarized text but, as we will show, we encounter several problems connected with the convoluted nature of the task, that is very different from the one reported in standard benchmarks of paraphrase detection and sentence similarity. Code and data available at https://github.com/roccotrip/protocols.
Reference annotated (or gold-standard) datasets are required for various common tasks such as training for machine learning systems or system validation. They are necessary to analyse or compare occurrences or items annotated by experts, or to compare objects resulting from any computational process to objects annotated by experts. But, even if reference annotated gold-standard corpora are required, their production is known as a difficult problem, from both a theoretical and practical point of view. Many studies devoted to theses issues conclude that multi-annotation is most of the time a necessity. That inter-annotator agreement measure, which is required to check the reliability of data and the reproducibility of an annotation task, and thus to establish a gold standard, is another thorny problem. Fine analysis of available metrics for this specific task then becomes essential. Our work is part of this effort and more precisely focuses on several problems, which are rarely discussed, although they are intrinsically linked with the interpretation of metrics. In particular, we focus here on the complex relations between agreement and reference (of which agreement among annotators is supposed to be an indicator), and the emergence of consensus. We also introduce the notion of consensuality as another relevant indicator.
Pretrained models through self-supervised learning have been recently introduced for both acoustic and language modeling. Applied to spoken language understanding tasks, these models have shown their great potential by improving the state-of-the-art performances on challenging benchmark datasets. In this paper, we present an error analysis reached by the use of such models on the French MEDIA benchmark dataset, known as being one of the most challenging benchmarks for the slot filling task among all the benchmarks accessible to the entire research community. One year ago, the state-of-art system reached a Concept Error Rate (CER) of 13.6% through the use of a end-to-end neural architecture. Some months later, a cascade approach based on the sequential use of a fine-tuned wav2vec2.0 model and a fine-tuned BERT model reaches a CER of 11.2%. This significant improvement raises questions about the type of errors that remain difficult to treat, but also about those that have been corrected using these models pre-trained through self-supervision learning on a large amount of data. This study brings some answers in order to better understand the limits of such models and open new perspectives to continue improving the performance.
To develop high-performance natural language understanding (NLU) models, it is necessary to have a benchmark to evaluate and analyze NLU ability from various perspectives. While the English NLU benchmark, GLUE, has been the forerunner, benchmarks are now being released for languages other than English, such as CLUE for Chinese and FLUE for French; but there is no such benchmark for Japanese. We build a Japanese NLU benchmark, JGLUE, from scratch without translation to measure the general NLU ability in Japanese. We hope that JGLUE will facilitate NLU research in Japanese.
A popular idea in Computer Assisted Language Learning (CALL) is to use multimodal annotated texts, with annotations typically including embedded audio and translations, to support L2 learning through reading. An important question is how to create good quality audio, which can be done either through human recording or by a Text-To-Speech (TTS) engine. We may reasonably expect TTS to be quicker and easier, but human to be of higher quality. Here, we report a study using the open source LARA platform and ten languages. Samples of audio totalling about five minutes, representing the same four passages taken from LARA versions of Saint-Exupèry’s “Le petit prince”, were provided for each language in both human and TTS form; the passages were chosen to instantiate the 2x2 cross product of the conditions dialogue, not-dialogue and humour, not-humour. 251 subjects used a web form to compare human and TTS versions of each item and rate the voices as a whole. For the three languages where TTS did best, English, French and Irish, the evidence from this study and the previous one it extended suggest that TTS audio is now pedagogically adequate and roughly comparable with a non-professional human voice in terms of exemplifying correct pronunciation and prosody. It was however still judged substantially less natural and less pleasant to listen to. No clear evidence was found to support the hypothesis that dialogue and humour pose special problems for TTS. All data and software will be made freely available.
A limited amount of studies investigates the role of model-agnostic adversarial behavior in toxic content classification. As toxicity classifiers predominantly rely on lexical cues, (deliberately) creative and evolving language-use can be detrimental to the utility of current corpora and state-of-the-art models when they are deployed for content moderation. The less training data is available, the more vulnerable models might become. This study is, to our knowledge, the first to investigate the effect of adversarial behavior and augmentation for cyberbullying detection. We demonstrate that model-agnostic lexical substitutions significantly hurt classifier performance. Moreover, when these perturbed samples are used for augmentation, we show models become robust against word-level perturbations at a slight trade-off in overall task performance. Augmentations proposed in prior work on toxicity prove to be less effective. Our results underline the need for such evaluations in online harm areas with small corpora.
The generation of referring expressions (REs) is a non-deterministic task. However, the algorithms for the generation of REs are standardly evaluated against corpora of written texts which include only one RE per each reference. Our goal in this work is firstly to reproduce one of the few studies taking the distributional nature of the RE generation into account. We add to this work, by introducing a method for exploring variation in human RE choice on the basis of longitudinal corpora - substantial corpora with a single human judgement (in the process of composition) per RE. We focus on the prediction of RE types, proper name, description and pronoun. We compare evaluations made against distributions over these types with evaluations made against parallel human judgements. Our results show agreement in the evaluation of learning algorithms against distributions constructed from parallel human evaluations and from longitudinal data.
Data-driven systems need to be evaluated to establish trust in the scientific approach and its applicability. In particular, this is true for Knowledge Graph (KG) Question Answering (QA), where complex data structures are made accessible via natural-language interfaces. Evaluating the capabilities of these systems has been a driver for the community for more than ten years while establishing different KGQA benchmark datasets. However, comparing different approaches is cumbersome. The lack of existing and curated leaderboards leads to a missing global view over the research field and could inject mistrust into the results. In particular, the latest and most-used datasets in the KGQA community, LC-QuAD and QALD, miss providing central and up-to-date points of trust. In this paper, we survey and analyze a wide range of evaluation results with significant coverage of 100 publications and 98 systems from the last decade. We provide a new central and open leaderboard for any KGQA benchmark dataset as a focal point for the community - https://kgqa.github.io/leaderboard/. Our analysis highlights existing problems during the evaluation of KGQA systems. Thus, we will point to possible improvements for future evaluations.
We present a multi-task learning framework for cross-lingual abstractive summarization to augment training data. Recent studies constructed pseudo cross-lingual abstractive summarization data to train their neural encoder-decoders. Meanwhile, we introduce existing genuine data such as translation pairs and monolingual abstractive summarization data into training. Our proposed method, Transum, attaches a special token to the beginning of the input sentence to indicate the target task. The special token enables us to incorporate the genuine data into the training data easily. The experimental results show that Transum achieves better performance than the model trained with only pseudo cross-lingual summarization data. In addition, we achieve the top ROUGE score on Chinese-English and Arabic-English abstractive summarization. Moreover, Transum also has a positive effect on machine translation. Experimental results indicate that Transum improves the performance from the strong baseline, Transformer, in Chinese-English, Arabic-English, and English-Japanese translation datasets.
This paper analyses how much context span is necessary to solve different context-related issues, namely, reference, ellipsis, gender, number, lexical ambiguity, and terminology when translating from English into Portuguese. We use the DELA corpus, which consists of 60 documents and six different domains (subtitles, literary, news, reviews, medical, and legislation). We find that the shortest context span to disambiguate issues can appear in different positions in the document including preceding, following, global, world knowledge. Moreover, the average length depends on the issue types as well as the domain. Moreover, we show that the standard approach of relying on only two preceding sentences as context might not be enough depending on the domain and issue types.
Document-level Neural Machine Translation aims to increase the quality of neural translation models by taking into account contextual information. Properly modelling information beyond the sentence level can result in improved machine translation output in terms of coherence, cohesion and consistency. Suitable corpora for context-level modelling are necessary to both train and evaluate context-aware systems, but are still relatively scarce. In this work we describe TANDO, a document-level corpus for the under-resourced Basque-Spanish language pair, which we share with the scientific community. The corpus is composed of parallel data from three different domains and has been prepared with context-level information. Additionally, the corpus includes contrastive test sets for fine-grained evaluations of gender and register contextual phenomena on both source and target language sides. To establish the usefulness of the corpus, we trained and evaluated baseline Transformer models and context-aware variants based on context concatenation. Our results indicate that the corpus is suitable for fine-grained evaluation of document-level machine translation systems.
In this work, we present the work that has been carried on in the MT4All CEF project and the resources that it has generated by leveraging recent research carried out in the field of unsupervised learning. In the course of the project 18 monolingual corpora for specific domains and languages have been collected, and 12 bilingual dictionaries and translation models have been generated. As part of the research, the unsupervised MT methodology based only on monolingual corpora (Artetxe et al., 2017) has been tested on a variety of languages and domains. Results show that in specialised domains, when there is enough monolingual in-domain data, unsupervised results are comparable to those of general domain supervised translation, and that, at any rate, unsupervised techniques can be used to boost results whenever very little data is available.
This paper introduces a multi-lingual database containing translated texts of COVID-19 mythbusters. The database has translations into 115 languages as well as the original English texts, of which the original texts are published by World Health Organization (WHO). This paper then presents preliminary analyses on latin-alphabet-based texts to see the potential of the database as a resource for multilingual linguistic analyses. The analyses on latin-alphabet-based texts gave interesting insights into the resource. While the amount of translated texts in each language was small, character bi-grams with normalization (lowercasing and removal of diacritics) was turned out to be an effective proxy for measuring the similarity of the languages, and the affinity ranking of language pairs could be obtained. Additionally, the hierarchical clustering analysis is performed using the character bigram overlap ratio of every possible pair of languages. The result shows the cluster of Germanic languages, Romance languages, and Southern Bantu languages. In sum, the multilingual database not only offers fixed set of materials in numerous languages, but also serves as a preliminary tool to identify the language family using text-based similarity measure of bigram overlap ratio.
Generative Pre-trained Transformers (GPTs) have recently been scaled to unprecedented sizes in the history of machine learning. These models, solely trained on the language modeling objective, have been shown to exhibit outstanding zero, one, and few-shot learning capabilities in a number of different tasks. Nevertheless, aside from anecdotal experiences, little is known regarding their multilingual capabilities, given the fact that the pre-training corpus is almost entirely composed of English text. In this work, we investigate its potential and limits in three tasks: extractive question-answering, text summarization and natural language generation for five different languages, as well as the effect of scale in terms of model size. Our results show that GPT-3 can be almost as useful for many languages as it is for English, with room for improvement if optimization of the tokenization is addressed.
Subtitles appear on screen as short pieces of text, segmented based on formal constraints (length) and syntactic/semantic criteria. Subtitle segmentation can be evaluated with sequence segmentation metrics against a human reference. However, standard segmentation metrics cannot be applied when systems generate outputs different than the reference, e.g. with end-to-end subtitling systems. In this paper, we study ways to conduct reference-based evaluations of segmentation accuracy irrespective of the textual content. We first conduct a systematic analysis of existing metrics for evaluating subtitle segmentation. We then introduce Sigma, a Subtitle Segmentation Score derived from an approximate upper-bound of BLEU on segmentation boundaries, which allows us to disentangle the effect of good segmentation from text quality. To compare Sigma with existing metrics, we further propose a boundary projection method from imperfect hypotheses to the true reference. Results show that all metrics are able to reward high quality output but for similar outputs system ranking depends on each metric’s sensitivity to error type. Our thorough analyses suggest Sigma is a promising segmentation candidate but its reliability over other segmentation metrics remains to be validated through correlations with human judgements.
Despite impressive progress in machine translation in recent years, it has occasionally been argued that current systems are still mainly based on pattern recognition and that further progress may be possible by using text understanding techniques, thereby e.g. looking at semantics of the type “Who is doing what to whom?”. In the current research we aim to take a small step into this direction. Assuming that semantic role labeling (SRL) grasps some of the relevant semantics, we automatically annotate the source language side of a standard parallel corpus, namely Europarl, with semantic roles. We then train a neural machine translation (NMT) system using the annotated corpus on the source language side, and the original unannotated corpus on the target language side. New text to be translated is first annotated by the same SRL system and then fed into the translation system. We compare the results to those of a baseline NMT system trained with unannotated text on both sides and find that the SRL-based system yields small improvements in terms of BLEU scores for each of the four language pairs under investigation, involving English, French, German, Greek and Spanish.
Natural Language Inference (NLI), also known as Recognizing Textual Entailment (RTE), has been one of the central tasks in Artificial Intelligence (AI) and Natural Language Processing (NLP). RTE between the two pieces of texts is a crucial problem, and it adds further challenges when involving two different languages, i.e., in the cross-lingual scenario. This paper proposes an effective transfer learning approach for cross-lingual NLI. We perform experiments on English-Hindi language pairs in the cross-lingual setting to find out that our novel loss formulation could enhance the performance of the baseline model by up to 2%. To assess the effectiveness of our method further, we perform additional experiments on every possible language pair using four European languages, namely French, German, Bulgarian, and Turkish, on top of XNLI dataset. Evaluation results yield up to 10% performance improvement over the respective baseline models, in some cases surpassing the state-of-the-art (SOTA). It is also to be noted that our proposed model has 110M parameters which is much lesser than the SOTA model having 220M parameters. Finally, we argue that our transfer learning-based loss objective is model agnostic and thus can be used with other deep learning-based architectures for cross-lingual NLI.
Specialist high-quality information is typically first available in English, and it is written in a language that may be difficult to understand by most readers. While Machine Translation technologies contribute to mitigate the first issue, the translated content will most likely still contain complex language. In order to investigate and address both problems simultaneously, we introduce Simple TICO-19, a new language resource containing manual simplifications of the English and Spanish portions of the TICO-19 corpus for Machine Translation of COVID-19 literature. We provide an in-depth description of the annotation process, which entailed designing an annotation manual and employing four annotators (two native English speakers and two native Spanish speakers) who simplified over 6,000 sentences from the English and Spanish portions of the TICO-19 corpus. We report several statistics on the new dataset, focusing on analysing the improvements in readability from the original texts to their simplified versions. In addition, we propose baseline methodologies for automatically generating the simplifications, translations and joint translation and simplifications contained in our dataset.
Recent work has demonstrated the importance of dealing with Multi-Word Terms (MWTs) in several Natural Language Processing applications. In particular, MWTs pose serious challenges for alignment and machine translation systems because of their syntactic and semantic properties. Thus, developing algorithms that handle MWTs is becoming essential for many NLP tasks. However, the availability of bilingual and more generally multi-lingual resources is limited, especially for low-resourced languages and in specialized domains. In this paper, we propose an approach for building comparable corpora and bilingual term dictionaries that help evaluate bilingual term alignment in comparable corpora. To that aim, we exploit parallel corpora to perform automatic bilingual MWT extraction and comparable corpus construction. Parallel information helps to align bilingual MWTs and makes it easier to build comparable specialized sub-corpora. Experimental validation on an existing dataset and on manually annotated data shows the interest of the proposed methodology.
This paper examines machine bias in language technology. Machine bias can affect machine learning algorithms when language models trained on large corpora include biased human decisions or reflect historical or social inequities, e.g. regarding gender and race. The focus of the paper is on gender bias in machine translation and we discuss a study conducted on Icelandic translations in the translation systems Google Translate and Vélþýðing.is. The results show a pattern which corresponds to certain societal ideas about gender. For example it seems to depend on the meaning of adjectives referring to people whether they appear in the masculine or feminine form. Adjectives describing positive personality traits were more likely to appear in masculine gender whereas the negative ones frequently appear in feminine gender. However, the opposite applied to appearance related adjectives. These findings unequivocally demonstrate the importance of being vigilant towards technology so as not to maintain societal inequalities and outdated views — especially in today’s digital world.
This paper presents an analysis of how dialogue act sequences vary across different datasets in order to anticipate the potential degradation in the performance of learned models during domain adaptation. We hypothesize the following: 1) dialogue sequences from related domains will exhibit similar n-gram frequency distributions 2) this similarity can be expressed by measuring the average Hamming distance between subsequences drawn from different datasets. Our experiments confirm that when dialogue acts sequences from two datasets are dissimilar they lie further away in embedding space, making it possible to train a classifier to discriminate between them even when the datasets are corrupted with noise. We present results from eight different datasets: SwDA, AMI (DialSum), GitHub, Hate Speech, Teams, Diplomacy Betrayal, SAMsum, and Military (Army). Our datasets were collected from many types of human communication including strategic planning, informal discussion, and social media exchanges. Our methodology provides intuition on the generalizability of dialogue models trained on different datasets. Based on our analysis, it is problematic to assume that machine learning models trained on one type of discourse will generalize well to other settings, due to contextual differences.
Interview is an efficient way to elicit knowledge from experts of different domains. In this paper, we introduce CIDC, an interview dialogue corpus in the culinary domain in which interviewers play an active role to elicit culinary knowledge from the cooking expert. The corpus consists of 308 interview dialogues (each about 13 minutes in length), which add up to a total of 69,000 utterances. We use a video conferencing tool for data collection, which allows us to obtain the facial expressions of the interlocutors as well as the screen-sharing contents. To understand the impact of the interlocutors’ skill level, we divide the experts into “semi-professionals’” and “enthusiasts” and the interviewers into “skilled interviewers” and “unskilled interviewers.” For quantitative analysis, we report the statistics and the results of the post-interview questionnaire. We also conduct qualitative analysis on the collected interview dialogues and summarize the salient patterns of how interviewers elicit knowledge from the experts. The corpus serves the purpose to facilitate future research on the knowledge elicitation mechanism in interview dialogues.
In this paper, we introduce a carefully designed and collected language resource: UgChDial – a Uyghur dialogue corpus based on a chatroom environment. The Uyghur Chat-based Dialogue Corpus (UgChDial) is divided into two parts: (1). Two-party dialogues and (2). Multi-party dialogues. We ran a series of 25, 120-minutes each, two-party chat sessions, totaling 7323 turns and 1581 question-response pairs. We created 16 different scenarios and topics to gather these two-party conversations. The multi-party conversations were compiled from chitchats in general channels as well as free chats in topic-oriented public channels, yielding 5588 unique turns and 838 question-response pairs. The initial purpose of this corpus is to study query-response pairs in Uyghur, building on an existing fine-grained response space taxonomy for English. We provide here initial annotation results on the Uyghur response space classification task using UgChDial.
This study investigates how the grounding process is composed and explores new interaction approaches that adapt to human cognitive processes that have not yet been significantly studied. The results of an experiment indicate that grounding through dialogue is mutually accepted among participants through holistic expressions and suggest that common ground among participants may not necessarily be formed in a bottom-up way through analytic expressions. These findings raise the possibility of a promising new approach to creating a human-like dialogue system that may be more suitable for natural human communication.
The main objective of this work is the elaboration and public release of BaSCo, the first corpus with annotated linguistic resources encompassing Basque-Spanish code-switching. The mixture of Basque and Spanish languages within the same utterance is popularly referred to as Euskañol, a widespread phenomenon among bilingual speakers in the Basque Country. Thus, this corpus has been created to meet the demand of annotated linguistic resources in Euskañol in research areas such as multilingual dialogue systems. The presented resource is the result of translating to Euskañol a compilation of texts in Basque and Spanish that were used for training the Natural Language Understanding (NLU) models of several task-oriented bilingual chatbots. Those chatbots were meant to answer specific questions associated with the administration, fiscal, and transport domains. In addition, they had the transverse potential to answer to greetings, requests for help, and chit-chat questions asked to chatbots. BaSCo is a compendium of 1377 tagged utterances with every sample annotated at three levels: (i) NLU semantic labels, considering intents and entities, (ii) code-switching proportion, and (iii) domain of origin.
Robots will eventually enter our daily lives and assist with a variety of tasks. Especially in the household domain, robots may become indispensable helpers by overtaking tedious tasks, e.g. keeping the place tidy. Their effectiveness and efficiency, however, depend on their ability to adapt to our needs, routines, and personal characteristics. Otherwise, they may not be accepted and trusted in our private domain. For enabling adaptation, the interaction between a human and a robot needs to be personalized. Therefore, the robot needs to collect personal information from the user. However, it is unclear how such sensitive data can be collected in an understandable way without losing a user’s trust in the system. In this paper, we present a conversational approach for explicitly collecting personal user information using natural dialogue. For creating a sound interactive personalization, we have developed an empathy-augmented dialogue strategy. In an online study, the empathy-augmented strategy was compared to a baseline dialogue strategy for interactive personalization. We have found the empathy-augmented strategy to perform notably friendlier. Overall, using dialogue for interactive personalization has generally shown positive user reception.
Taking minutes is an essential component of every meeting, although the goals, style, and procedure of this activity (“minuting” for short) can vary. Minuting is a rather unstructured writing activity and is affected by who is taking the minutes and for whom the intended minutes are. With the rise of online meetings, automatic minuting would be an important benefit for the meeting participants as well as for those who might have missed the meeting. However, automatically generating meeting minutes is a challenging problem due to a variety of factors including the quality of automatic speech recorders (ASRs), availability of public meeting data, subjective knowledge of the minuter, etc. In this work, we present the first of its kind dataset on Automatic Minuting. We develop a dataset of English and Czech technical project meetings which consists of transcripts generated from ASRs, manually corrected, and minuted by several annotators. Our dataset, AutoMin, consists of 113 (English) and 53 (Czech) meetings, covering more than 160 hours of meeting content. Upon acceptance, we will publicly release (aaa.bbb.ccc) the dataset as a set of meeting transcripts and minutes, excluding the recordings for privacy reasons. A unique feature of our dataset is that most meetings are equipped with more than one minute, each created independently. Our corpus thus allows studying differences in what people find important while taking the minutes. We also provide baseline experiments for the community to explore this novel problem further. To the best of our knowledge AutoMin is probably the first resource on minuting in English and also in a language other than English (Czech).
Age-related stereotypes are pervasive in our society, and yet have been under-studied in the NLP community. Here, we present a method for extracting age-related stereotypes from Twitter data, generating a corpus of 300,000 over-generalizations about four contemporary generations (baby boomers, generation X, millennials, and generation Z), as well as “old” and “young” people more generally. By employing word-association metrics, semi-supervised topic modelling, and density-based clustering, we uncover many common stereotypes as reported in the media and in the psychological literature, as well as some more novel findings. We also observe trends consistent with the existing literature, namely that definitions of “young” and “old” age appear to be context-dependent, stereotypes for different generations vary across different topics (e.g., work versus family life), and some age-based stereotypes are distinct from generational stereotypes. The method easily extends to other social group labels, and therefore can be used in future work to study stereotypes of different social categories. By better understanding how stereotypes are formed and spread, and by tracking emerging stereotypes, we hope to eventually develop mitigating measures against such biased statements.
We present a new corpus of Twitter data annotated for codeswitching and borrowing between Spanish and English. The corpus contains 9,500 tweets annotated at the token level with codeswitches, borrowings, and named entities. This corpus differs from prior corpora of codeswitching in that we attempt to clearly define and annotate the boundary between codeswitching and borrowing and do not treat common “internet-speak” (lol, etc.) as codeswitching when used in an otherwise monolingual context. The result is a corpus that enables the study and modeling of Spanish-English borrowing and codeswitching on Twitter in one dataset. We present baseline scores for modeling the labels of this corpus using Transformer-based language models. The annotation itself is released with a CC BY 4.0 license, while the text it applies to is distributed in compliance with the Twitter terms of service.
Mental disorders are a serious and increasingly relevant public health issue. NLP methods have the potential to assist with automatic mental health disorder detection, but building annotated datasets for this task can be challenging; moreover, annotated data is very scarce for disorders other than depression. Understanding the commonalities between certain disorders is also important for clinicians who face the problem of shifting standards of diagnosis. We propose that transfer learning with linguistic features can be useful for approaching both the technical problem of improving mental disorder detection in the context of data scarcity, and the clinical problem of understanding the overlapping symptoms between certain disorders. In this paper, we target four disorders: depression, PTSD, anorexia and self-harm. We explore multi-aspect transfer learning for detecting mental disorders from social media texts, using deep learning models with multi-aspect representations of language (including multiple types of interpretable linguistic features). We explore different transfer learning strategies for cross-disorder and cross-platform transfer, and show that transfer learning can be effective for improving prediction performance for disorders where little annotated data is available. We offer insights into which linguistic features are the most useful vehicles for transferring knowledge, through ablation experiments, as well as error analysis.
The emergence of the COVID-19 pandemic and the first global infodemic have changed our lives in many different ways. We relied on social media to get the latest information about COVID-19 pandemic and at the same time to disseminate information. The content in social media consisted not only health related advice, plans, and informative news from policymakers, but also contains conspiracies and rumors. It became important to identify such information as soon as they are posted to make an actionable decision (e.g., debunking rumors, or taking certain measures for traveling). To address this challenge, we develop and publicly release the first largest manually annotated Arabic tweet dataset, ArCovidVac, for COVID-19 vaccination campaign, covering many countries in the Arab region. The dataset is enriched with different layers of annotation, including, (i) Informativeness more vs. less importance of the tweets); (ii) fine-grained tweet content types (e.g., advice, rumors, restriction, authenticate news/information); and (iii) stance towards vaccination (pro-vaccination, neutral, anti-vaccination). Further, we performed in-depth analysis of the data, exploring the popularity of different vaccines, trending hashtags, topics, and presence of offensiveness in the tweets. We studied the data for individual types of tweets and temporal changes in stance towards vaccine. We benchmarked the ArCovidVac dataset using transformer architectures for informativeness, content types, and stance detection.
Proactively identifying misinformation spreaders is an important step towards mitigating the impact of fake news on our society. In this paper, we introduce a new contemporary Reddit dataset for fake news spreader analysis, called FACTOID, monitoring political discussions on Reddit since the beginning of 2020. The dataset contains over 4K users with 3.4M Reddit posts, and includes, beyond the users’ binary labels, also their fine-grained credibility level (very low to very high) and their political bias strength (extreme right to extreme left). As far as we are aware, this is the first fake news spreader dataset that simultaneously captures both the long-term context of users’ historical posts and the interactions between them. To create the first benchmark on our data, we provide methods for identifying misinformation spreaders by utilizing the social connections between the users along with their psycho-linguistic features. We show that the users’ social interactions can, on their own, indicate misinformation spreading, while the psycho-linguistic features are mostly informative in non-neural classification settings. In a qualitative analysis we observe that detecting affective mental processes correlates negatively with right-biased users, and that the openness to experience factor is lower for those who spread fake news.
Anglicisms are a challenge in German speech recognition. Due to their irregular pronunciation compared to native German words, automatically generated pronunciation dictionaries often contain incorrect phoneme sequences for Anglicisms. In this work, we propose a multitask sequence-to-sequence approach for grapheme-to-phoneme conversion to improve the phonetization of Anglicisms. We extended a grapheme-to-phoneme model with a classification task to distinguish Anglicisms from native German words. With this approach, the model learns to generate different pronunciations depending on the classification result. We used our model to create supplementary Anglicism pronunciation dictionaries to be added to an existing German speech recognition model. Tested on a special Anglicism evaluation set, we improved the recognition of Anglicisms compared to a baseline model, reducing the word error rate by a relative 1 % and the Anglicism error rate by a relative 3 %. With our experiment, we show that multitask learning can help solving the challenge of Anglicisms in German speech recognition.
We present SDS-200, a corpus of Swiss German dialectal speech with Standard German text translations, annotated with dialect, age, and gender information of the speakers. The dataset allows for training speech translation, dialect recognition, and speech synthesis systems, among others. The data was collected using a web recording tool that is open to the public. Each participant was given a text in Standard German and asked to translate it to their Swiss German dialect before recording it. To increase the corpus quality, recordings were validated by other participants. The data consists of 200 hours of speech by around 4000 different speakers and covers a large part of the Swiss German dialect landscape. We release SDS-200 alongside a baseline speech translation model, which achieves a word error rate (WER) of 30.3 and a BLEU score of 53.1 on the SDS-200 test set. Furthermore, we use SDS-200 to fine-tune a pre-trained XLS-R model, achieving 21.6 WER and 64.0 BLEU.
This paper builds upon recent work in leveraging the corpora and tools originally used to develop speech technologies for corpus-based linguistic studies. We address the non-canonical realization of consonants in connected speech and we focus on voicing alternation phenomena of stops in 5 standard varieties of Romance languages (French, Italian, Spanish, Portuguese, Romanian). For these languages, both large scale corpora and speech recognition systems were available for the study. We use forced alignment with pronunciation variants and machine learning techniques to examine to what extent such frequent phenomena characterize languages and what are the most triggering factors. The results confirm that voicing alternations occur in all Romance languages. Automatic classification underlines that surrounding contexts and segment duration are recurring contributing factors for modeling voicing alternation. The results of this study also demonstrate the new role that machine learning techniques such as classification algorithms can play in helping to extract linguistic knowledge from speech and to suggest interesting research directions.
Our main goal is to study the interactions between speakers according to their gender and role in broadcast media. In this paper, we propose an extensive study of gender and overlap annotations in various speech corpora mainly dedicated to diarisation or transcription tasks. We point out the issue of the heterogeneity of the annotation guidelines for both overlapping speech and gender categories. On top of that, we analyse how the speech content (casual speech, meetings, debate, interviews, etc.) impacts the distribution of overlapping speech segments. On a small dataset of 93 recordings from LCP French channel, we intend to characterise the interactions between speakers according to their gender. Finally, we propose a method which aims to highlight active speech areas in terms of interactions between speakers. Such a visualisation tool could improve the efficiency of qualitative studies conducted by researchers in human sciences.
This paper presents a semi-automatic approach to create a diachronic corpus of voices balanced for speaker’s age, gender, and recording period, according to 32 categories (2 genders, 4 age ranges and 4 recording periods). Corpora were selected at French National Institute of Audiovisual (INA) to obtain at least 30 speakers per category (a total of 960 speakers; only 874 have be found yet). For each speaker, speech excerpts were extracted from audiovisual documents using an automatic pipeline consisting of speech detection, background music and overlapped speech removal and speaker diarization, used to present clean speaker segments to human annotators identifying target speakers. This pipeline proved highly effective, cutting down manual processing by a factor of ten. Evaluation of the quality of the automatic processing and of the final output is provided. It shows the automatic processing compare to up-to-date process, and that the output provides high quality speech for most of the selected excerpts. This method is thus recommendable for creating large corpora of known target speakers.
We present DiscoGeM, a crowdsourced corpus of 6,505 implicit discourse relations from three genres: political speech, literature, and encyclopedic texts. Each instance was annotated by 10 crowd workers. Various label aggregation methods were explored to evaluate how to obtain a label that best captures the meaning inferred by the crowd annotators. The results show that a significant proportion of discourse relations in DiscoGeM are ambiguous and can express multiple relation senses. Probability distribution labels better capture these interpretations than single labels. Further, the results emphasize that text genre crucially affects the distribution of discourse relations, suggesting that genre should be included as a factor in automatic relation classification. We make available the newly created DiscoGeM corpus, as well as the dataset with all annotator-level labels. Both the corpus and the dataset can facilitate a multitude of applications and research purposes, for example to function as training data to improve the performance of automatic discourse relation parsers, as well as facilitate research into non-connective signals of discourse relations.
Broadcast political debate is a core pillar of democracy: it is the public’s easiest access to opinions that shape policies and enables the general public to make informed choices. With QT30, we present the largest corpus of analysed dialogical argumentation ever created (19,842 utterances, 280,000 words) and also the largest corpus of analysed broadcast political debate to date, using 30 episodes of BBC’s ‘Question Time’ from 2020 and 2021. Question Time is the prime institution in UK broadcast political debate and features questions from the public on current political issues, which are responded to by a weekly panel of five figures of UK politics and society. QT30 is highly argumentative and combines language of well-versed political rhetoric with direct, often combative, justification-seeking of the general public. QT30 is annotated with Inference Anchoring Theory, a framework well-known in argument mining, which encodes the way arguments and conflicts are created and reacted to in dialogical settings. The resource is freely available at http://corpora.aifdb.org/qt30.
The empirical quantification of the quality of a contribution to a political discussion is at the heart of deliberative theory, the subdiscipline of political science which investigates decision-making in deliberative democracy. Existing annotation on deliberative quality is time-consuming and carried out by experts, typically resulting in small datasets which also suffer from strong class imbalance. Scaling up such annotations with automatic tools is desirable, but very challenging. We take up this challenge and explore different strategies to improve the prediction of deliberative quality dimensions (justification, common good, interactivity, respect) in a standard dataset. Our results show that simple data augmentation techniques successfully alleviate data imbalance. Classifiers based on linguistic features (textual complexity and sentiment/polarity) and classifiers integrating argument quality annotations (from the argument mining community in NLP) were consistently outperformed by transformer-based models, with or without data augmentation.
Natural language inherently consists of implicit and underspecified phrases, which represent potential sources of misunderstanding. In this paper, we present a data set of such phrases in English from instructional texts together with multiple possible clarifications. Our data set, henceforth called CLAIRE, is based on a corpus of revision histories from wikiHow, from which we extract human clarifications that resolve an implicit or underspecified phrase. We show how language modeling can be used to generate alternate clarifications, which may or may not be compatible with the human clarification. Based on plausibility judgements for each clarification, we define the task of distinguishing between plausible and implausible clarifications. We provide several baseline models for this task and analyze to what extent different clarifications represent multiple readings as a first step to investigate misunderstandings caused by implicit/underspecified language in instructional texts.
The paper presents a multilingual database aimed to be used as a tool for typological analysis of response constructions called discourse formulae (DF), cf. English ‘No way¡ or French ‘Ça va¡ ( ‘all right’). The two primary qualities that make DF of theoretical interest for linguists are their idiomaticity and the special nature of their meanings (cf. consent, refusal, negation), determined by their dialogical function. The formal and semantic structures of these items are language-specific. Compiling a database with DF from various languages would help estimate the diversity of DF in both of these aspects, and, at the same time, establish some frequently occurring patterns. The DF in the database are accompanied with glosses and assigned with multiple tags, such as pragmatic function, additional semantics, the illocutionary type of the context, etc. As a starting point, Russian, Serbian and Slovene DF are included into the database. This data already shows substantial grammatical and lexical variability.
In this paper we present the Serbian part of the ELTeC multilingual corpus of novels written in the time period 1840-1920. The corpus is being built in order to test various distant reading methods and tools with the aim of re-thinking the European literary history. We present the various steps that led to the production of the Serbian sub-collection: the novel selection and retrieval, text preparation, structural annotation, POS-tagging, lemmatization and named entity recognition. The Serbian sub-collection was published on different platforms in order to make it freely available to various users. Several use examples show that this sub-collection is usefull for both close and distant reading approaches.
Automatizing the process of understanding the global narrative structure of long texts and stories is still a major challenge for state-of-the-art natural language understanding systems, particularly because annotated data is scarce and existing annotation workflows do not scale well to the annotation of complex narrative phenomena. In this work, we focus on the identification of narrative levels in texts corresponding to stories that are embedded in stories. Lacking sufficient pre-annotated training data, we explore a solution to deal with data scarcity that is common in machine learning: the automatic augmentation of an existing small data set of annotated samples with the help of data synthesis. We present a workflow for narrative level detection, that includes the operationalization of the task, a model, and a data augmentation protocol for automatically generating narrative texts annotated with breaks between narrative levels. Our experiments suggest that narrative levels in long text constitute a challenging phenomenon for state-of-the-art NLP models, but generating training data synthetically does improve the prediction results considerably.
Spelling normalisation is a useful step in the study and analysis of historical language texts, whether it is manual analysis by experts or automatic analysis using downstream natural language processing (NLP) tools. Not only does it help to homogenise the variable spelling that often exists in historical texts, but it also facilitates the use of off-the-shelf contemporary NLP tools, if contemporary spelling conventions are used for normalisation. We present FREEMnorm, a new benchmark for the normalisation of Early Modern French (from the 17th century) into contemporary French and provide a thorough comparison of three different normalisation methods: ABA, an alignment-based approach and MT-approaches, (both statistical and neural), including extensive parameter searching, which is often missing in the normalisation literature.
anguage models for historical states of language are becoming increasingly important to allow the optimal digitisation and analysis of old textual sources. Because these historical states are at the same time more complex to process and more scarce in the corpora available, this paper presents recent efforts to overcome this difficult situation. These efforts include producing a corpus, creating the model, and evaluating it with an NLP task currently used by scholars in other ongoing projects.
Identifying the high level structure of texts provides important information when performing distant reading analysis. The structure of texts is not necessarily linear, as transitions, such as changes in the scenery or flashbacks, can be present. As a first step in identifying this structure, we aim to identify transitions in texts. Previous work (Heyns and van Zaanen, 2021) proposed a system that can successfully identify one transition in literary texts. The text is split in snippets and LDA is applied, resulting in a sequence of topics. A transition is introduced at the point that separates the topics (before and after the point) best. In this article, we extend the existing system such that it can detect multiple transitions. Additionally, we introduce a new system that inherently handles multiple transitions in texts. The new system also relies on LDA information, but is more robust than the previous system. We apply these systems to texts with known transitions (as they are constructed by concatenating text snippets stemming from different source texts) and evaluation both systems on texts with one transition and texts with two transitions. As both systems rely on LDA to identify transitions between snippets, we also show the impact of varying the number of LDA topics on the results as well. The new system consistently outperforms the previous system, not only on texts with multiple transitions, but also on single boundary texts.
Parliamentary transcripts provide a valuable resource to understand the reality and know about the most important facts that occur over time in our societies. Furthermore, the political debates captured in these transcripts facilitate research on political discourse from a computational social science perspective. In this paper we release the first version of a newly compiled corpus from Basque parliamentary transcripts. The corpus is characterized by heavy Basque-Spanish code-switching, and represents an interesting resource to study political discourse in contrasting languages such as Basque and Spanish. We enrich the corpus with metadata related to relevant attributes of the speakers and speeches (language, gender, party...) and process the text to obtain named entities and lemmas. The obtained metadata is then used to perform a detailed corpus analysis which provides interesting insights about the language use of the Basque political representatives across time, parties and gender.
Although studied for several decades, the syntactic properties of experiencer-object (EO) verbs are still under discussion, while most analyses are not supported by substantial corpus data. With GerEO, we intend to fill this lacuna for German EO-verbs by presenting a large-scale database of more than 10,000 examples for 64 verbs (up to 200 per verb) from a newspaper corpus annotated for several syntactic and semantic features relevant for their analysis, including the overall syntactic construction, the semantic stimulus type, and the form of a possible stimulus preposition, i.e. a preposition heading a PP that indicates (a part/aspect of) the stimulus. Non-psych occurrences of the verbs are not excluded from the database but marked as such to make a comparison possible. Data of this kind can be used to develop and test theoretical hypotheses on the properties of EO-verbs, aid in the construction of experiments as well as provide training and test data for AI systems.
Classifying citations according to their purpose and importance is a challenging task that has gained considerable interest in recent years. This interest has been primarily driven by the need to create more transparent, efficient, merit-based reward systems in academia; a system that goes beyond simple bibliometric measures and considers the semantics of citations. Such systems that quantify and classify the influence of citations can act as edges that link knowledge nodes to a graph and enable efficient knowledge discovery. While a number of researchers have experimented with a variety of models, these experiments are typically limited to single-domain applications and the resulting models are hardly comparable. Recently, two Citation Context Classification (3C) shared tasks (at WOSP2020 and SDP2021) created the first benchmark enabling direct comparison of citation classification approaches, revealing the crucial impact of supplementary data on the performance of models. Reflecting from the findings of these shared tasks, we are releasing a new multi-disciplinary dataset, ACT2, an extended SDP 3C shared task dataset. This modified corpus has annotations for both citation function and importance classes newly enriched with supplementary contextual and non-contextual feature sets the selection of which follows from the lists of features used by the more successful teams in these shared tasks. Additionally, we include contextual features for cited papers (e.g. Abstract of the cited paper), which most existing datasets lack, but which have a lot of potential to improve results. We describe the methodology used for feature extraction and the challenges involved in the process. The feature enriched ACT2 dataset is available at https://github.com/oacore/ACT2.
This paper describes the continuation of a project that aims at establishing an interoperable annotation schema for quantification phenomena as part of the ISO suite of standards for semantic annotation, known as the Semantic Annotation Framework. After a break, caused by the Covid-19 pandemic, the project was relaunched in early 2022 with a second working draft of an annotation scheme, which is discussed in this paper. Keywords: semantic annotation, quantification, interoperability, annotation schema, ISO standard
We present the Hindi-Telugu Parallel Corpus of different technical domains such as Natural Science, Computer Science, Law and Healthcare along with the General domain. The qualitative corpus consists of 700K parallel sentences of which 535K sentences were created using multiple methods such as extract, align and review of Hindi-Telugu corpora, end-to-end human translation, iterative back-translation driven post-editing and around 165K parallel sentences were collected from available sources in the public domain. We present the comparative assessment of created parallel corpora for representativeness and diversity. The corpus has been pre-processed for machine translation, and we trained a neural machine translation system using it and report state-of-the-art baseline results on the developed development set over multiple domains and on available benchmarks. With this, we define a new task on Domain Machine Translation for low resource language pairs such as Hindi and Telugu. The developed corpus (535K) is freely available for non-commercial research and to the best of our knowledge, this is the well curated, largest, publicly available domain parallel corpus for Hindi-Telugu.
This paper introduces a new Magahi-Hindi-English (MHE) code-mixed data-set for similar language identification (SMLID), where Magahi is a less-resourced minority language. This corpus provides a language id at two levels: word and sentence. This data-set is the first Magahi-Hindi-English code-mixed data-set for similar language identification task. Furthermore, we will discuss the complexity of the data-set and provide a few baselines for the language identification task.
We introduce a dataset built around a large collection of TV (and movie) series. Those are filled with challenging multi-party dialogues. Moreover, TV series come with a very active fan base that allows the collection of metadata and accelerates annotation. With 16 TV and movie series, Bazinga! amounts to 400+ hours of speech and 8M+ tokens, including 500K+ tokens annotated with the speaker, addressee, and entity linking information. Along with the dataset, we also provide a baseline for speaker diarization, punctuation restoration, and person entity recognition. The results demonstrate the difficulty of the tasks and of transfer learning from models trained on mono-speaker audio or written text, which is more widely available. This work is a step towards better multi-party dialogue structuring and understanding. Bazinga! is available at hf.co/bazinga. Because (a large) part of Bazinga! is only partially annotated, we also expect this dataset to foster research towards self- or weakly-supervised learning methods.
In this paper, we present the Ellogon Web Annotation Tool. It is a collaborative, web-based annotation tool built upon the Ellogon infrastructure offering an improved user experience and adaptability to various annotation scenarios by making good use of the latest design practices and web development frameworks. Being in development for many years, this paper describes its current architecture, along with the recent modifications that extend the existing functionalities and the new features that were added. The new version of the tool offers document analytics, annotation inspection and comparison features, a modern UI, and formatted text import (e.g. TEI XML documents, rendered with simple markup). We present two use cases that serve as two examples of different annotation scenarios to demonstrate the new functionalities. An appropriate (user-supplied, XML-based) annotation schema is used for each scenario. The first schema contains the relevant components for representing concepts, moral values, and ideas. The second includes all the necessary elements for annotating argumentative units in a document and their binary relations.
The WeCanTalk (WCT) Corpus is a new multi-language, multi-modal resource for speaker recognition. The corpus contains Cantonese, Mandarin and English telephony and video speech data from over 200 multilingual speakers located in Hong Kong. Each speaker contributed at least 10 telephone conversations of 8-10 minutes’ duration collected via a custom telephone platform based in Hong Kong. Speakers also uploaded at least 3 videos in which they were both speaking and visible, along with one selfie image. At least half of the calls and videos for each speaker were in Cantonese, while their remaining recordings featured one or more different languages. Both calls and videos were made in a variety of noise conditions. All speech and video recordings were audited by experienced multilingual annotators for quality including presence of the expected language and for speaker identity. The WeCanTalk Corpus has been used to support the NIST 2021 Speaker Recognition Evaluation and will be published in the LDC catalog.
This paper describes an approach aiming at utilizing Wiktionary data for creating specialized lexical datasets which can be used for enriching other lexical (semantic) resources or for generating datasets that can be used for evaluating or improving NLP tasks, like Word Sense Disambiguation, Word-in-Context challenges, or Sense Linking across lexicons and dictionaries. We have focused on Wiktionary data about pronunciation information in English, and grammatical number and grammatical gender in German.
Formal documents often are organized into sections of text, each with a title, and extracting this structure remains an under-explored aspect of natural language processing. This iterative title-text structure is valuable data for building models for headline generation and section title generation, but there is no corpus that contains web documents annotated with titles and prose texts. Therefore, we propose the first title-text dataset on web documents that incorporates a wide variety of domains to facilitate downstream training. We also introduce STAPI (Section Title And Prose text Identifier), a two-step system for labeling section titles and prose text in HTML documents. To filter out unrelated content like document footers, its first step involves a filter that reads HTML documents and proposes a set of textual candidates. In the second step, a typographic classifier takes the candidates from the filter and categorizes each one into one of the three pre-defined classes (title, prose text, and miscellany). We show that STAPI significantly outperforms two baseline models in terms of title-text identification. We release our dataset along with a web application to facilitate supervised and semi-supervised training in this domain.
ELTE Poetry Corpus is a database that stores canonical Hungarian poetry with automatically generated annotations of the poems’ structural units, grammatical features and sound devices, i.e. rhyme patterns, rhyme pairs, rhythm, alliterations and the main phonological features of words. The corpus has an open access online query tool with several search functions. The paper presents the main stages of the annotation process and the tools used for each stage. The TEI XML format of the different versions of the corpus, each of which contains an increasing number of annotation layers, is presented as well. We have also specified our own XML format for the corpus, slightly different from TEI, in order to make it easier and faster to execute queries on the corpus. We discuss the results of a manual evaluation of the quality of automatic annotation of rhythm, as well as the results of an automatic evaluation of different rule sets used for the automatic annotation of rhyme patterns. Finally, the paper gives an overview of the main functions of the online query tool developed for the corpus.
Word Problem Solving remains a challenging and interesting task in NLP. A lot of research has been carried out to solve different genres of word problems with various complexity levels in recent years. However, most of the publicly available datasets and work has been carried out for English. Recently there has been a surge in this area of word problem solving in Chinese with the creation of large benchmark datastes. Apart from these two languages, labeled benchmark datasets for low resource languages are very scarce. This is the first attempt to address this issue for any Indian Language, especially Hindi. In this paper, we present HAWP (Hindi Arithmetic Word Problems), a dataset consisting of 2336 arithmetic word problems in Hindi. We also developed baseline systems for solving these word problems. We also propose a new evaluation technique for word problem solvers taking equation equivalence into account.
The paper describes the Bulgarian Event Corpus (BEC). The annotation scheme is based on CIDOC-CRM ontology and on the English Framenet, adjusted for our task. It includes two main layers: named entities and events with their roles. The corpus is multi-domain and mainly oriented towards Social Sciences and Humanities (SSH). It will be used for: extracting knowledge and making it available through the Bulgaria-centric Knowledge Graph; further developing an annotation scheme that handles multiple domains in SSH; training automatic modules for the most important knowledge-based tasks, such as domain-specific and nested NER, NEL, event detection and profiling. Initial experiments were conducted on standard NER task due to complexity of the dataset and the rich NE annotation scheme. The results are promising with respect to some labels and give insights on handling better other ones. These experiments serve also as error detection modules that would help us in scheme re-design. They are a basis for further and more complex tasks, such as nested NER, NEL and event detection.
The Story Cloze Test (SCT) is designed for training and evaluating machine learning algorithms for narrative understanding and inferences. The SOTA models can achieve over 90% accuracy on predicting the last sentence. However, it has been shown that high accuracy can be achieved by merely using surface-level features. We suspect these models may not truly understand the story. Based on the SCT dataset, we constructed a human-labeled and human-verified commonsense knowledge inference dataset. Given the first four sentences of a story, we asked crowd-source workers to choose from four types of narrative inference for deciding the ending sentence and which sentence contributes most to the inference. We accumulated data on 1871 stories, and three human workers labeled each story. Analysis of the intra-category and inter-category agreements show a high level of consensus. We present two new tasks for predicting the narrative inference categories and contributing sentences. Our results show that transformer-based models can reach SOTA performance on the original SCT task using transfer learning but don’t perform well on these new and more challenging tasks.
We present GTP-SW3, a 3.5 billion parameter autoregressive language model, trained on a newly created 100 GB Swedish corpus. This paper provides insights with regards to data collection and training, while highlights the challenges of proper model evaluation. The results of quantitive evaluation through perplexity indicate that GPT-SW3 is a competent model in comparison with existing autoregressive models of similar size. Additionally, we perform an extensive prompting study which reveals the good text generation capabilities of GTP-SW3.
This paper describes a system for interactive poem generation, which combines neural language models (LMs) for poem generation with explicit constraints that can be set by users on form, topic, emotion, and rhyming scheme. LMs cannot learn such constraints from the data, which is scarce with respect to their needs even for a well-resourced language such as French. We propose a method to generate verses and stanzas by combining LMs with rule-based algorithms, and compare several approaches for adjusting the words of a poem to a desired combination of topics or emotions. An approach to automatic rhyme setting using a phonetic dictionary is proposed as well. Our system has been demonstrated at public events, and log analysis shows that users found it engaging.
Online trolls increase social costs and cause psychological damage to individuals. With the proliferation of automated accounts making use of bots for trolling, it is difficult for targeted individual users to handle the situation both quantitatively and qualitatively. To address this issue, we focus on automating the method to counter trolls, as counter responses to combat trolls encourage community users to maintain ongoing discussion without compromising freedom of expression. For this purpose, we propose a novel dataset for automatic counter response generation. In particular, we constructed a pair-wise dataset that includes troll comments and counter responses with labeled response strategies, which enables models fine-tuned on our dataset to generate responses by varying counter responses according to the specified strategy. We conducted three tasks to assess the effectiveness of our dataset and evaluated the results through both automatic and human evaluation. In human evaluation, we demonstrate that the model fine-tuned with our dataset shows a significantly improved performance in strategy-controlled sentence generation.
Numerical tables are widely employed to communicate or report the classification performance of machine learning (ML) models with respect to a set of evaluation metrics. For non-experts, domain knowledge is required to fully understand and interpret the information presented by numerical tables. This paper proposes a new natural language generation (NLG) task where neural models are trained to generate textual explanations, analytically describing the classification performance of ML models based on the metrics’ scores reported in the tables. Presenting the generated texts along with the numerical tables will allow for a better understanding of the classification performance of ML models. We constructed a dataset comprising numerical tables paired with their corresponding textual explanations written by experts to facilitate this NLG task. Experiments on the dataset are conducted by fine-tuning pre-trained language models (T5 and BART) to generate analytical textual explanations conditioned on the information in the tables. Furthermore, we propose a neural module, Metrics Processing Unit (MPU), to improve the performance of the baselines in terms of correctly verbalising the information in the corresponding table. Evaluation and analysis conducted indicate, that exploring pre-trained models for data-to-text generation leads to better generalisation performance and can produce high-quality textual explanations.
We present Barch, a new English dataset of human-written summaries describing bar charts. This dataset contains 47 charts based on a selection of 18 topics. Each chart is associated with one of the four intended messages expressed in the chart title. Using crowdsourcing, we collected around 20 summaries per chart, or one thousand in total. The text of the summaries is aligned with the chart data as well as with analytical inferences about the data drawn by humans. Our datasets is one of the first to explore the effect of intended messages on the data descriptions in chart summaries. Additionally, it lends itself well to the task of training data-driven systems for chart-to-text generation. We provide results on the performance of state-of-the-art neural generation models trained on this dataset and discuss the strengths and shortcomings of different models.
We tackle the problem of neural headline generation in a low-resource setting, where only limited amount of data is available to train a model. We compare the ideal high-resource scenario on English with results obtained on a smaller subset of the same data and also run experiments on two small news corpora covering low-resource languages, Croatian and Estonian. Two options for headline generation in a multilingual low-resource scenario are investigated: a pretrained multilingual encoder-decoder model and a combination of two pretrained language models, one used as an encoder and the other as a decoder, connected with a cross-attention layer that needs to be trained from scratch. The results show that the first approach outperforms the second one by a large margin. We explore several data augmentation and pretraining strategies in order to improve the performance of both models and show that while we can drastically improve the second approach using these strategies, they have little to no effect on the performance of the pretrained encoder-decoder model. Finally, we propose two new measures for evaluating the performance of the models besides the classic ROUGE scores.
Pre-trained language models have established the state-of-the-art on various natural language processing tasks, including dialogue summarization, which allows the reader to quickly access key information from long conversations in meetings, interviews or phone calls. However, such dialogues are still difficult to handle with current models because the spontaneity of the language involves expressions that are rarely present in the corpora used for pre-training the language models. Moreover, the vast majority of the work accomplished in this field has been focused on English. In this work, we present a study on the summarization of spontaneous oral dialogues in French using several language specific pre-trained models: BARThez, and BelGPT-2, as well as multilingual pre-trained models: mBART, mBARThez, and mT5. Experiments were performed on the DECODA (Call Center) dialogue corpus whose task is to generate abstractive synopses from call center conversations between a caller and one or several agents depending on the situation. Results show that the BARThez models offer the best performance far above the previous state-of-the-art on DECODA. We further discuss the limits of such pre-trained models and the challenges that must be addressed for summarizing spontaneous dialogues.
Lexical Simplification is the process of reducing the lexical complexity of a text by replacing difficult words with easier to read (or understand) expressions while preserving the original information and meaning. In this paper we introduce ALEXSIS, a new dataset for this task, and we use ALEXSIS to benchmark Lexical Simplification systems in Spanish. The paper describes the evaluation of three kind of approaches to Lexical Simplification, a thesaurus-based approach, a single transformers-based approach, and a combination of transformers. We also report state of the art results on a previous Lexical Simplification dataset for Spanish.
IARPA’s Better Extraction from Text Towards Enhanced Retrieval (BETTER) Program created multiple multilingual datasets to spawn and evaluate cross-language information extraction and information retrieval research and development in zero-shot conditions. The first set of these resources for information extraction, the “Abstract” data will be released to the public at LREC 2022 in four languages to champion further information extraction work in this area. This paper presents the event and argument annotation in the Abstract Evaluation phase of BETTER, as well as the data collection, preparation, partitioning and mark-up of the datasets.
Named Entity Recognition (NER) is an important task in information extraction. However, due to the lack of labelled corpora, biomedical NER has scarcely been studied in Vietnamese compared to English. To address this situation, we have constructed VietBioNER, a labelled NER corpus of Vietnamese academic biomedical text. The corpus focuses specifically on supporting tuberculosis surveillance, and was constructed by collecting scientific papers and grey literature related to tuberculosis symptoms and diagnostics. We manually annotated a small set of the collected documents with five categories of named entities: Organisation, Location, Date and Time, Symptom and Disease, and Diagnostic Procedure. Inter-annotator agreement ranges from 70.59% and 95.89% F-score according to entity category. In this paper, we make available two splits of the corpus, corresponding to traditional supervised learning and few-shot learning settings. We also provide baseline results for both of these settings, in addition to a dictionary-based approach, as a means to stimulate further research into Vietnamese biomedical NER. Although supervised methods produce results that are far superior to the other two approaches, the fact that even one-shot learning can outperform the dictionary-based method provides evidence that further research into few-shot learning on this text type would be worthwhile.
Forced labour is the most common type of modern slavery, and it is increasingly gaining the attention of the research and social community. Recent studies suggest that artificial intelligence (AI) holds immense potential for augmenting anti-slavery action. However, AI tools need to be developed transparently in cooperation with different stakeholders. Such tools are contingent on the availability and access to domain-specific data, which are scarce due to the near-invisible nature of forced labour. To the best of our knowledge, this paper presents the first openly accessible English corpus annotated for multi-class and multi-label forced labour detection. The corpus consists of 989 news articles retrieved from specialised data sources and annotated according to risk indicators defined by the International Labour Organization (ILO). Each news article was annotated for two aspects: (1) indicators of forced labour as classification labels and (2) snippets of the text that justify labelling decisions. We hope that our data set can help promote research on explainability for multi-class and multi-label text classification. In this work, we explain our process for collecting the data underpinning the proposed corpus, describe our annotation guidelines and present some statistical analysis of its content. Finally, we summarise the results of baseline experiments based on different variants of the Bidirectional Encoder Representation from Transformer (BERT) model.
This paper presents Wojood, a corpus for Arabic nested Named Entity Recognition (NER). Nested entities occur when one entity mention is embedded inside another entity mention. Wojood consists of about 550K Modern Standard Arabic (MSA) and dialect tokens that are manually annotated with 21 entity types including person, organization, location, event and date. More importantly, the corpus is annotated with nested entities instead of the more common flat annotations. The data contains about 75K entities and 22.5% of which are nested. The inter-annotator evaluation of the corpus demonstrated a strong agreement with Cohen’s Kappa of 0.979 and an F1-score of 0.976. To validate our data, we used the corpus to train a nested NER model based on multi-task learning using the pre-trained AraBERT (Arabic BERT). The model achieved an overall micro F1-score of 0.884. Our corpus, the annotation guidelines, the source code and the pre-trained model are publicly available.
In this work, we present the first corpus for German Adverse Drug Reaction (ADR) detection in patient-generated content. The data consists of 4,169 binary annotated documents from a German patient forum, where users talk about health issues and get advice from medical doctors. As is common in social media data in this domain, the class labels of the corpus are very imbalanced. This and a high topic imbalance make it a very challenging dataset, since often, the same symptom can have several causes and is not always related to a medication intake. We aim to encourage further multi-lingual efforts in the domain of ADR detection and provide preliminary experiments for binary classification using different methods of zero- and few-shot learning based on a multi-lingual model. When fine-tuning XLM-RoBERTa first on English patient forum data and then on the new German data, we achieve an F1-score of 37.52 for the positive class. We make the dataset and models publicly available for the community.
Despite remarkable advances in the development of language resources over the recent years, there is still a shortage of annotated, publicly available corpora covering (German) medical language. With the initial release of the German Guideline Program in Oncology NLP Corpus (GGPONC), we have demonstrated how such corpora can be built upon clinical guidelines, a widely available resource in many natural languages with a reasonable coverage of medical terminology. In this work, we describe a major new release for GGPONC. The corpus has been substantially extended in size and re-annotated with a new annotation scheme based on SNOMED CT top level hierarchies, reaching high inter-annotator agreement (γ=.94). Moreover, we annotated elliptical coordinated noun phrases and their resolutions, a common language phenomenon in (not only German) scientific documents. We also trained BERT-based named entity recognition models on this new data set, which achieve high performance on short, coarse-grained entity spans (F1=.89), while the rate of boundary errors increases for long entity spans. GGPONC is freely available through a data use agreement. The trained named entity recognition models, as well as the detailed annotation guide, are also made publicly available.
This paper presents ClinIDMap, a tool for mapping identifiers between clinical ontologies and lexical resources. ClinIDMap interlinks identifiers from UMLS, SMOMED-CT, ICD-10 and the corresponding Wikipedia articles for concepts from the UMLS Metathesaurus. Our main goal is to provide semantic interoperability across the clinical concepts from various knowledge bases. As a side effect, the mapping enriches already annotated corpora in multiple languages with new labels. For instance, spans manually annotated with IDs from UMLS can be annotated with Semantic Types and Groups, and its corresponding SNOMED CT and ICD-10 IDs. We also experiment with sequence labelling models for detecting Diagnosis and Procedures concepts and for detecting UMLS Semantic Groups trained on Spanish, English, and bilingual corpora obtained with the new mapping procedure. The ClinIDMap tool is publicly available.
In our paper, we present a novel corpus of historical legal documents on the Romanian public procurement legislation and an annotated subset of draft bills that have been screened by legal experts and identified as impacting past public procurement legislation. Using the manual annotations provided by the experts, we attempt to automatically identify future draft bills that have the potential to impact existing policies on public procurement.
The impressive progress in NLP techniques has been driven by the development of multi-task benchmarks such as GLUE and SuperGLUE. While these benchmarks focus on tasks for one or two input sentences, there has been exciting work in designing efficient techniques for processing much longer inputs. In this paper, we present MuLD: a new long document benchmark consisting of only documents over 10,000 tokens. By modifying existing NLP tasks, we create a diverse benchmark which requires models to successfully model long-term dependencies in the text. We evaluate how existing models perform, and find that our benchmark is much more challenging than their ‘short document’ equivalents. Furthermore, by evaluating both regular and efficient transformers, we show that models with increased context length are better able to solve the tasks presented, suggesting that future improvements in these models are vital for solving similar long document problems. We release the data and code for baselines to encourage further research on efficient NLP models.
This paper proposes a new cross-document coreference resolution (CDCR) dataset for identifying co-referring radiological findings and medical devices across a patient’s radiology reports. Our annotated corpus contains 5872 mentions (findings and devices) spanning 638 MIMIC-III radiology reports across 60 patients, covering multiple imaging modalities and anatomies. There are a total of 2292 mention chains. We describe the annotation process in detail, highlighting the complexities involved in creating a sizable and realistic dataset for radiology CDCR. We apply two baseline methods–string matching and transformer language models (BERT)–to identify cross-report coreferences. Our results indicate the requirement of further model development targeting better understanding of domain language and context to address this challenging and unexplored task. This dataset can serve as a resource to develop more advanced natural language processing CDCR methods in the future. This is one of the first attempts focusing on CDCR in the clinical domain and holds potential in benefiting physicians and clinical research through long-term tracking of radiology findings.
The business world has changed due to the 21st century economy, where borders have melted and trades became free. Nowadays,competition is no longer only at the local market level but also at the global level. In this context, the World Wide Web has become a major source of information for companies and professionals to keep track of their complex, rapidly changing, and competitive business environment. A lot of effort is nonetheless needed to collect and analyze this information due to information overload problem and the huge number of web pages to process and analyze. In this paper, we propose the BizRel resource, the first multilingual (French,English, Spanish, and Chinese) dataset for automatic extraction of binary business relations involving organizations from the web.This dataset is used to train several monolingual and cross-lingual deep learning models to detect these relations in texts. Our results are encouraging, demonstrating the effectiveness of such a resource for both research and business communities. In particular, we believe multilingual business relation extraction systems are crucial tools for decision makers to identify links between specific market stakeholders and build business networks which enable to anticipate changes and discover new threats or opportunities. Our work is therefore an important direction toward such tools.
With their Discovery of Inference Rules from Text (DIRT) algorithm, Lin and Pantel (2001) made a seminal contribution to the field of rule acquisition from text, by adapting the distributional hypothesis of Harris (1954) to rules that model binary relations such as X treat Y. DIRT’s relevance is renewed in today’s neural era given the recent focus on interpretability in the field of natural language processing. We propose a novel take on the DIRT algorithm, where we implement the distributional hypothesis using the contextualized embeddings provided by BERT, a transformer-network-based language model (Vaswani et al. 2017; Devlin et al. 2018). In particular, we change the similarity measure between pairs of slots (i.e., the set of words matched by a rule) from the original formula that relies on lexical items to a formula computed using contextualized embeddings. We empirically demonstrate that this new similarity method yields a better implementation of the distributional hypothesis, and this, in turn, yields rules that outperform the original algorithm in the question answering-based evaluation proposed by Lin and Pantel (2001).
Unfortunately, offensive language in social media is a common phenomenon nowadays. It harms many people and vulnerable groups. Therefore, automated detection of offensive language is in high demand and it is a serious challenge in multilingual domains. Various machine learning approaches combined with natural language techniques have been applied for this task lately. This paper contributes to this area from several aspects: (1) it introduces a new dataset of annotated Facebook comments in Hebrew; (2) it describes a case study with multiple supervised models and text representations for a task of offensive language detection in three languages, including two Semitic (Hebrew and Arabic) languages; (3) it reports evaluation results of cross-lingual and multilingual learning for detection of offensive content in Semitic languages; and (4) it discusses the limitations of these settings.
In the field of Japanese medical information extraction, few analyzing tools are available and relation extraction is still an under-explored topic. In this paper, we first propose a novel relation annotation schema for investigating the medical and temporal relations between medical entities in Japanese medical reports. We experiment with the practical annotation scenarios by separately annotating two different types of reports. We design a pipeline system with three components for recognizing medical entities, classifying entity modalities, and extracting relations. The empirical results show accurate analyzing performance and suggest the satisfactory annotation quality, the superiority of the latest contextual embedding models. and the feasible annotation strategy for high-accuracy demand.
Modern approaches in Natural Language Processing (NLP) require, ideally, large amounts of labelled data for model training. However, new language resources, for example, for Named Entity Recognition (NER), Co-reference Resolution (CR), Entity Linking (EL) and Relation Extraction (RE), naming a few of the most popular tasks in NLP, have always been challenging to create since manual text annotations can be very time-consuming to acquire. While there may be an acceptable amount of labelled data available for some of these tasks in one language, there may be a lack of datasets in another. WEXEA is a tool to exhaustively annotate entities in the English Wikipedia. Guidelines for editors of Wikipedia articles result, on the one hand, in only a few annotations through hyperlinks, but on the other hand, make it easier to exhaustively annotate the rest of these articles with entities than starting from scratch. We propose the following main improvements to WEXEA: Creating multi-lingual corpora, improved entity annotations using a proven NER system, annotating dates and times. A brief evaluation of the annotation quality of WEXEA is added.
We present EpidBioBERT, a biosurveillance epidemiological document tagger for disease surveillance over PADI-Web system. Our model is trained on PADI-Web corpus which contains news articles on Animal Diseases Outbreak extracted from the web. We train a classifier to discriminate between relevant and irrelevant documents based on their epidemiological thematic feature content in preparation for further epidemiology information extraction. Our approach proposes a new way to perform epidemiological document classification by enriching epidemiological thematic features namely disease, host, location and date, which are used as inputs to our epidemiological document classifier. We adopt a pre-trained biomedical language model with a novel fine tuning approach that enriches these epidemiological thematic features. We find these thematic features rich enough to improve epidemiological document classification over a smaller data set than initially used in PADI-Web classifier. This improves the classifiers ability to avoid false positive alerts on disease surveillance systems. To further understand information encoded in EpidBioBERT, we experiment the impact of each epidemiology thematic feature on the classifier under ablation studies. We compare our biomedical pre-trained approach with a general language model based model finding that thematic feature embeddings pre-trained on general English documents are not rich enough for epidemiology classification task. Our model achieves an F1-score of 95.5% over an unseen test set, with an improvement of +5.5 points on F1-Score on the PADI-Web classifier with nearly half the training data set.
The de-identification of sensible data, also known as automatic textual anonymisation, is essential for data sharing and reuse, both for research and commercial purposes. The first step for data anonymisation is the detection of sensible entities. In this work, we present four new datasets for named entity detection in Spanish in the legal domain. These datasets have been generated in the framework of the MAPA project, three smaller datasets have been manually annotated and one large dataset has been automatically annotated, with an estimated error rate of around 14%. In order to assess the quality of the generated datasets, we have used them to fine-tune a battery of entity-detection models, using as foundation different pre-trained language models: one multilingual, two general-domain monolingual and one in-domain monolingual. We compare the results obtained, which validate the datasets as a valuable resource to fine-tune models for the task of named entity detection. We further explore the proposed methodology by applying it to a real use case scenario.
Recent advancements in natural language processing (NLP) have reshaped the industry, with powerful language models such as GPT-3 achieving superhuman performance on various tasks. However, the increasing complexity of such models turns them into “black boxes”, creating uncertainty about their internal operation and decision-making. Tsetlin Machine (TM) employs human-interpretable conjunctive clauses in propositional logic to solve complex pattern recognition problems and has demonstrated competitive performance in various NLP tasks. In this paper, we propose ConvTextTM, a novel convolutional TM architecture for text classification. While legacy TM solutions treat the whole text as a corpus-specific set-of-words (SOW), ConvTextTM breaks down the text into a sequence of text fragments. The convolution over the text fragments opens up for local position-aware analysis. Further, ConvTextTM eliminates the dependency on a corpus-specific vocabulary. Instead, it employs a generic SOW formed by the tokenization scheme of the Bidirectional Encoder Representations from Transformers (BERT). The convolution binds together the tokens, allowing ConvTextTM to address the out-of-vocabulary problem as well as spelling errors. We investigate the local explainability of our proposed method using clause-based features. Extensive experiments are conducted on seven datasets, to demonstrate that the accuracy of ConvTextTM is either superior or comparable to state-of-the-art baselines.
Comparative Question Answering (cQA) is the task of providing concrete and accurate responses to queries such as: “Is Lyft cheaper than a regular taxi?” or “What makes a mortgage different from a regular loan?”. In this paper, we propose two new open-domain real-world datasets for identifying and labeling comparative questions. While the first dataset contains instances of English questions labeled as comparative vs. non-comparative, the second dataset provides additional labels including the objects and the aspects of comparison. We conduct several experiments that evaluate the soundness of our datasets. The evaluation of our datasets using various classifiers show promising results that reach close-to-human results on a binary classification task with a neural model using ALBERT embeddings. When approaching the unsupervised sequence labeling task, some headroom remains.
Relation extraction is a core problem for natural language processing in the biomedical domain. Recent research on relation extraction showed that prompt-based learning improves the performance on both fine-tuning on full training set and few-shot training. However, less effort has been made on domain-specific tasks where good prompt design can be even harder. In this paper, we investigate prompting for biomedical relation extraction, with experiments on the ChemProt dataset. We present a simple yet effective method to systematically generate comprehensive prompts that reformulate the relation extraction task as a cloze-test task under a simple prompt formulation. In particular, we experiment with different ranking scores for prompt selection. With BioMed-RoBERTa-base, our results show that prompting-based fine-tuning obtains gains by 14.21 F1 over its regular fine-tuning baseline, and 1.14 F1 over SciFive-Large, the current state-of-the-art on ChemProt. Besides, we find prompt-based learning requires fewer training examples to make reasonable predictions. The results demonstrate the potential of our methods in such a domain-specific relation extraction task.
The growing interest in named entity recognition (NER) in various domains has led to the creation of different benchmark datasets, often with slightly different annotation guidelines. To better understand the different NER benchmark datasets for the domain of English literature and their impact on the evaluation of NER tools, we analyse two existing annotated datasets and create two additional gold standard datasets. Following on from this, we evaluate the performance of two NER tools, one domain-specific and one general-purpose NER tool, using the four gold standards, and analyse the sources for the differences in the measured performance. Our results show that the performance of the two tools varies significantly depending on the gold standard used for the individual evaluations.
There is an increasing need for the ability to model fine-grained opinion shifts of social media users, as concerns about the potential polarizing social effects increase. However, the lack of publicly available datasets that are suitable for the task presents a major challenge. In this paper, we introduce an innovative annotated dataset for modeling subtle opinion fluctuations and detecting fine-grained stances. The dataset includes a sufficient amount of stance polarity and intensity labels per user over time and within entire conversational threads, thus making subtle opinion fluctuations detectable both in long term and in short term. All posts are annotated by non-experts and a significant portion of the data is also annotated by experts. We provide a strategy for recruiting suitable non-experts. Our analysis of the inter-annotator agreements shows that the resulting annotations obtained from the majority vote of the non-experts are of comparable quality to the annotations of the experts. We provide analyses of the stance evolution in short term and long term levels, a comparison of language usage between users with vacillating and resolute attitudes, and fine-grained stance detection baselines.
Despite the large number of computational resources for emotion recognition, there is a lack of data sets relying on appraisal models. According to Appraisal theories, emotions are the outcome of a multi-dimensional evaluation of events. In this paper, we present APPReddit, the first corpus of non-experimental data annotated according to this theory. After describing its development, we compare our resource with enISEAR, a corpus of events created in an experimental setting and annotated for appraisal. Results show that the two corpora can be mapped notwithstanding different typologies of data and annotations schemes. A SVM model trained on APPReddit predicts four appraisal dimensions without significant loss. Merging both corpora in a single training set increases the prediction of 3 out of 4 dimensions. Such findings pave the way to a better performing classification model for appraisal prediction.
One of the challenges of aspect-based sentiment analysis is the implicit mention of aspects. These are more difficult to identify and may require world knowledge to do so. In this work, we evaluate frequency-based, hybrid, and machine learning methods, including the use of the pre-trained BERT language model, in the task of extracting aspect terms in opinionated texts in Portuguese, emphasizing the analysis of implicit aspects. Besides the comparative evaluation of methods, the differential of this work lies in the analysis’s novelty using a typology of implicit aspects that shows the knowledge needed to identify each implicit aspect term, thus allowing a mapping of the strengths and weaknesses of each method.
In recent years, AI research has demonstrated enormous potential for the benefit of humanity and society. While often better than its human counterparts in classification and pattern recognition tasks, however, AI still struggles with complex tasks that require commonsense reasoning such as natural language understanding. In this context, the key limitations of current AI models are: dependency, reproducibility, trustworthiness, interpretability, and explainability. In this work, we propose a commonsense-based neurosymbolic framework that aims to overcome these issues in the context of sentiment analysis. In particular, we employ unsupervised and reproducible subsymbolic techniques such as auto-regressive language models and kernel methods to build trustworthy symbolic representations that convert natural language to a sort of protolanguage and, hence, extract polarity from text in a completely interpretable and explainable manner.
In this paper, we launch a new Universal Dependencies treebank for an endangered language from Amazonia: Kakataibo, a Panoan language spoken in Peru. We first discuss the collaborative methodology implemented, which proved effective to create a treebank in the context of a Computational Linguistic course for undergraduates. Then, we describe the general details of the treebank and the language-specific considerations implemented for the proposed annotation. We finally conduct some experiments on part-of-speech tagging and syntactic dependency parsing. We focus on monolingual and transfer learning settings, where we study the impact of a Shipibo-Konibo treebank, another Panoan language resource.
Norwegian has been one of many languages lacking sufficient available text to train quality language models. In an attempt to bridge this gap, we introduce the Norwegian Colossal Corpus (NCC), which comprises 49GB of clean Norwegian textual data containing over 7B words. The NCC is composed of different and varied sources, ranging from books and newspapers to government documents and public reports, showcasing the various uses of the Norwegian language in society. The corpus contains mainly Norwegian Bokmål and Norwegian Nynorsk. Each document in the corpus is tagged with metadata that enables the creation of sub-corpora for specific needs. Its structure makes it easy to combine with large web archives that for licensing reasons could not be distributed together with the NCC. By releasing this corpus openly to the public, we hope to foster the creation of both better Norwegian language models and multilingual language models with support for Norwegian.
The paper presents novel resources and experiments for Buddhist Sanskrit, broadly defined here including all the varieties of Sanskrit in which Buddhist texts have been transmitted. We release a novel corpus of Buddhist texts, a novel corpus of general Sanskrit and word similarity and word analogy datasets for intrinsic evaluation of Buddhist Sanskrit embeddings models. We compare the performance of word2vec and fastText static embeddings models, with default and optimized parameter settings, as well as contextual models BERT and GPT-2, with different training regimes (including a transfer learning approach using the general Sanskrit corpus) and different embeddings construction regimes (given the encoder layers). The results show that for semantic similarity the fastText embeddings yield the best results, while for word analogy tasks BERT embeddings work the best. We also show that for contextual models the optimal layer combination for embedding construction is task dependant, and that pretraining the contextual embeddings models on a reference corpus of general Sanskrit is beneficial, which is a promising finding for future development of embeddings for less-resourced languages and domains.
This paper describes the process of data processing and training of an automatic speech recognition (ASR) system for Cook Islands Māori (CIM), an Indigenous language spoken by approximately 22,000 people in the South Pacific. We transcribed four hours of speech from adults and elderly speakers of the language and prepared two experiments. First, we trained three ASR systems: one statistical, Kaldi; and two based on Deep Learning, DeepSpeech and XLSR-Wav2Vec2. Wav2Vec2 tied with Kaldi for lowest character error rate (CER=6±1) and was slightly behind in word error rate (WER=23±2 versus WER=18±2 for Kaldi). This provides evidence that Deep Learning ASR systems are reaching the performance of statistical methods on small datasets, and that they can work effectively with extremely low-resource Indigenous languages like CIM. In the second experiment we used Wav2Vec2 to train models with held-out speakers. While the performance decreased (CER=15±7, WER=46±16), the system still showed considerable learning. We intend to use ASR to accelerate the documentation of CIM, using newly transcribed texts to improve the ASR and also generate teaching and language revitalization materials. The trained model is available under a license based on the Kaitiakitanga License, which provides for non-commercial use while retaining control of the model by the Indigenous community.
Protest events provide information about social and political conflicts, the state of social cohesion and democratic conflict management, as well as the state of civil society in general. Social scientists are therefore interested in the systematic observation of protest events. With this paper, we release the first German language resource of protest event related article excerpts published in local news outlets. We use this dataset to train and evaluate transformer-based text classifiers to automatically detect relevant newspaper articles. Our best approach reaches a binary F1-score of 93.3 %, which is a promising result for our goal to support political science research. However, in a second experiment, we show that our model does not generalize equally well when applied to data from time periods and localities other than our training sample. To make protest event detection more robust, we test two ways of alternative preprocessing. First, we find that letting the classifier concentrate on sentences around protest keywords improves the F1-score for out-of-sample data up to +4 percentage points. Second, against our initial intuition, masking of named entities during preprocessing does not improve the generalization in terms of F1-scores. However, it leads to a significantly improved recall of the models.
This paper presents text mining approaches on German-speaking job advertisements to enable social science research on the development of the labour market over the last 30 years. In order to build text mining applications providing information about profession and main task of a job, as well as experience and ICT skills needed, we experiment with transfer learning and domain adaptation. Our main contribution consists in building language models which are adapted to the domain of job advertisements, and their assessment on a broad range of machine learning problems. Our findings show the large value of domain adaptation in several respects. First, it boosts the performance of fine-tuned task-specific models consistently over all evaluation experiments. Second, it helps to mitigate rapid data shift over time in our special domain, and enhances the ability to learn from small updates with new, labeled task data. Third, domain-adaptation of language models is efficient: With continued in-domain pre-training we are able to outperform general-domain language models pre-trained on ten times more data. We share our domain-adapted language models and data with the research community.
Patronizing and Condescending Language (PCL) is a subtle but harmful type of discourse, yet the task of recognizing PCL remains under-studied by the NLP community. Recognizing PCL is challenging because of its subtle nature, because available datasets are limited in size, and because this task often relies on some form of commonsense knowledge. In this paper, we study to what extent PCL detection models can be improved by pre-training them on other, more established NLP tasks. We find that performance gains are indeed possible in this way, in particular when pre-training on tasks focusing on sentiment, harmful language and commonsense morality. In contrast, for tasks focusing on political speech and social justice, no or only very small improvements were witnessed. These findings improve our understanding of the nature of PCL.
This paper introduces HeLI-OTS, an off-the-shelf text language identification tool using the HeLI language identification method. The HeLI-OTS language identifier is equipped with language models for 200 languages and licensed for academic as well as commercial use. We present the HeLI method and its use in our previous research. Then we compare the performance of the HeLI-OTS language identifier with that of fastText on two different data sets, showing that fastText favors the recall of common languages, whereas HeLI-OTS reaches both high recall and high precision for all languages. While introducing existing off-the-shelf language identification tools, we also give a picture of digital humanities-related research that uses such tools. The validity of the results of such research depends on the results given by the language identifier used, and especially for research focusing on the less common languages, the tendency to favor widely used languages might be very detrimental, which Heli-OTS is now able to remedy.
Parallel corpora are ideal for extracting a multilingual named entity (MNE) resource, i.e., a dataset of names translated into multiple languages. Prior work on extracting MNE datasets from parallel corpora required resources such as large monolingual corpora or word aligners that are unavailable or perform poorly for underresourced languages. We present CLC-BN, a new method for creating an MNE resource, and apply it to the Parallel Bible Corpus, a corpus of more than 1000 languages. CLC-BN learns a neural transliteration model from parallel-corpus statistics, without requiring any other bilingual resources, word aligners, or seed data. Experimental results show that CLC-BN clearly outperforms prior work. We release an MNE resource for 1340 languages and demonstrate its effectiveness in two downstream tasks: knowledge graph augmentation and bilingual lexicon induction.
In this paper we will discuss our preliminary work towards the construction of a WordNet for Old English, taking our inspiration from other similar WN construction projects for ancient languages such as Ancient Greek, Latin and Sanskrit. The Old English WordNet (OldEWN) will build upon this innovative work in a number of different ways which we articulate in the article, most importantly by treateating figurative meaning as a ‘first-class citizen’ in the structuring of the semantic system. From a more practical perspective we will describe our plan to utilize a pre-existing lexicographic resource and the naisc system to automatically compile a provisional version of the WordNet which will then be checked and enriched by Old English experts.
This paper presents PFN-DE, a new, parsing- and annotation-oriented framenet for German, with almost 15,000 frames, covering 11,300 verb lemmas. The resource was developed in the context of a Danish/German social-media study on hate speech and has a strong focus on coverage, robustness and cross-language comparability. A simple annotation scheme for argument roles meshes directly with the output of a syntactic parser, facilitating frame disambiguation through slot-filler conditions based on valency, syntactic function and semantic noun class. We discuss design principles for the framenet and the frame tagger using it, and present statistics for frame and role distribution at both the lexicon (type) and corpus (token) levels. In an evaluation run on Twitter data, the parser-based frame annotator achieved an overall F-score for frame senses of 93.6%.
Robot-Assisted minimally invasive robotic surgery is the gold standard for the surgical treatment of many pathological conditions, and several manuals and academic papers describe how to perform these interventions. These high-quality, often peer-reviewed texts are the main study resource for medical personnel and consequently contain essential procedural domain-specific knowledge. The procedural knowledge therein described could be extracted, e.g., on the basis of semantic parsing models, and used to develop clinical decision support systems or even automation methods for some procedure’s steps. However, natural language understanding algorithms such as, for instance, semantic role labelers have lower efficacy and coverage issues when applied to domain others than those they are typically trained on (i.e., newswire text). To overcome this problem, starting from PropBank frames, we propose a new linguistic resource specific to the robotic-surgery domain, named Robotic Surgery Procedural Framebank (RSPF). We extract from robotic-surgical texts verbs and nouns that describe surgical actions and extend PropBank frames by adding any of new lemmas, frames or role sets required to cover missing lemmas, specific frames describing the surgical significance, or new semantic roles used in procedural surgical language. Our resource is publicly available and can be used to annotate corpora in the surgical domain to train and evaluate Semantic Role Labeling (SRL) systems in a challenging fine-grained domain setting.
Understanding child language development requires accurately representing children’s lexicons. However, much of the past work modeling children’s vocabulary development has utilized adult-based measures. The present investigation asks whether using corpora that captures the language input of young children more accurately represents children’s vocabulary knowledge. We present a newly-created toddler corpus that incorporates transcripts of child-directed conversations, the text of picture books written for preschoolers, and dialog from G-rated movies to approximate the language input a North American preschooler might hear. We evaluate the utility of the new corpus for modeling children’s vocabulary development by building and analyzing different semantic network models and comparing them to norms based on vocabulary norms for toddlers in this age range. More specifically, the relations between words in our semantic networks were derived from skip-gram neural networks (Word2Vec) trained on our toddler corpus or on Google news. Results revealed that the models built from the toddler corpus were more accurate at predicting toddler vocabulary growth than the adult-based corpus. These results speak to the importance of selecting a corpus that matches the population of interest.
We apply Formal Concept Analysis (FCA) to organize and to improve the quality of Démonette2, a French derivational database, through a detection of both missing and spurious derivations in the database. We represent each derivational family as a graph. Given that the subgraph relation exists among derivational families, FCA can group families and represent them in a partially ordered set (poset). This poset is also useful for improving the database. A family is regarded as a possible anomaly (meaning that it may have missing and/or spurious derivations) if its derivational graph is almost, but not completely identical to a large number of other families.
We present the current status of a new ontology for representing constitutive elements of Sign Languages (SL). This development emerged from investigations on how to represent multimodal lexical data in the OntoLex-Lemon framework, with the goal to publish such data in the Linguistic Linked Open Data (LLOD) cloud. While studying the literature and various sites dealing with sign languages, we saw the need to harmonise all the data categories (or features) defined and used in those sources, and to organise them in an ontology to which lexical descriptions in OntoLex-Lemon could be linked. We make the code of the first version of this ontology available, so that it can be further developed collaboratively by both the Linked Data and the SL communities
A commonsense knowledge resource organizes common sense that is not necessarily correct all the time, but most people are expected to know or believe. Such knowledge resources have recently been actively built and utilized in artificial intelligence, particularly natural language processing. In this paper, we discuss an important but not significantly discussed the issue of semantic gaps potentially existing in a commonsense knowledge graph and propose a machine learning-based approach to detect a semantic gap that may inhibit the proper chaining of knowledge triples. In order to establish this line of research, we created a pilot dataset from ConceptNet, in which chains consisting of two adjacent triples are sampled, and the validity of each chain is human-annotated. We also devised a few baseline methods for detecting the semantic gaps and compared them in small-scale experiments. Although the experimental results suggest that the detection of semantic gaps may not be a trivial task, we achieved several insights to further push this research direction, including the potential efficacy of sense embeddings and contextualized word representations enabled by a pre-trained language model.
We present Semi-Structured Explanations for COPA (COPA-SSE), a new crowdsourced dataset of 9,747 semi-structured, English common sense explanations for Choice of Plausible Alternatives (COPA) questions. The explanations are formatted as a set of triple-like common sense statements with ConceptNet relations but freely written concepts. This semi-structured format strikes a balance between the high quality but low coverage of structured data and the lower quality but high coverage of free-form crowdsourcing. Each explanation also includes a set of human-given quality ratings. With their familiar format, the explanations are geared towards commonsense reasoners operating on knowledge graphs and serve as a starting point for ongoing work on improving such systems. The dataset is available at https://github.com/a-brassard/copa-sse.
GRhOOT, the German RhetOrical OnTology, is a domain ontology of 110 rhetorical figures in the German language. The overall goal of building an ontology of rhetorical figures in German is not only the formal representation of different rhetorical figures, but also allowing for their easier detection, thus improving sentiment analysis, argument mining, detection of hate speech and fake news, machine translation, and many other tasks in which recognition of non-literal language plays an important role. The challenge of building such ontologies lies in classifying the figures and assigning adequate characteristics to group them, while considering their distinctive features. The ontology of rhetorical figures in the Serbian language was used as a basis for our work. Besides transferring and extending the concepts of the Serbian ontology, we ensured completeness and consistency by using description logic and SPARQL queries. Furthermore, we show a decision tree to identify figures and suggest a usage scenario on how the ontology can be utilized to collect and annotate data.
Large-scale diachronic corpus studies covering longer time periods are difficult if more than one corpus are to be consulted and, as a result, different formats and annotation schemas need to be processed and queried in a uniform, comparable and replicable manner. We describes the application of the Flexible Integrated Transformation and Annotation eNgineering (Fintan) platform for studying word order in German using syntactically annotated corpora that represent its entire written history. Focusing on nominal dative and accusative arguments, this study hints at two major phases in the development of scrambling in modern German. Against more recent assumptions, it supports the traditional view that word order flexibility decreased over time, but it also indicates that this was a relatively sharp transition in Early New High German. The successful case study demonstrates the potential of Fintan and the underlying LLOD technology for historical linguistics, linguistic typology and corpus linguistics. The technological contribution of this paper is to demonstrate the applicability of Fintan for querying across heterogeneously annotated corpora, as previously, it had only been applied for transformation tasks. With its focus on quantitative analysis, Fintan is a natural complement for existing multi-layer technologies that focus on query and exploration.
Although the Universal Dependencies initiative today allows for cross-linguistically consistent annotation of morphology and syntax in treebanks for several languages, syntactically annotated corpora are not yet interoperable with many lexical resources that describe properties of the words that occur therein. In order to cope with such limitation, we propose to adopt the principles of the Linguistic Linked Open Data community, to describe and publish dependency treebanks as LLOD. In particular, this paper illustrates the approach pursued in the LiLa Knowledge Base, which enables interoperability between corpora and lexical resources for Latin, to publish as Linguistic Linked Open Data the annotation layers of two versions of a Medieval Latin treebank (the Index Thomisticus Treebank).
Olfactory references play a crucial role in our memory and, more generally, in our experiences, since researchers have shown that smell is the sense that is most directly connected with emotions. Nevertheless, only few works in NLP have tried to capture this sensory dimension from a computational perspective. One of the main challenges is the lack of a systematic and consistent taxonomy of olfactory information, where concepts are organised also in a multi-lingual perspective. WordNet represents a valuable starting point in this direction, which can be semi-automatically extended taking advantage of Google n-grams and of existing language models. In this work we describe the process that has led to the semi-automatic development of a taxonomy for olfactory information in four languages (English, French, German and Italian), detailing the different steps and the intermediate evaluations. Along with being multi-lingual, the taxonomy also encloses temporal marks for olfactory terms thus making it a valuable resource for historical content analysis. The resource has been released and is freely available.
Today, natural language processing heavily relies on pre-trained large language models. Even though such models are criticized for the poor interpretability, they still yield state-of-the-art solutions for a wide set of very different tasks. While lots of probing studies have been conducted to measure the models’ awareness of grammatical knowledge, semantic probing is less popular. In this work, we introduce the probing pipeline to study the representedness of semantic relations in transformer language models. We show that in this task, attention scores are nearly as expressive as the layers’ output activations, despite their lesser ability to represent surface cues. This supports the hypothesis that attention mechanisms are focusing not only on the syntactic relational information but also on the semantic one.
Recently, many studies have focused on developing dialogue systems that enable collaborative work; however, they rarely focus on creative tasks. Collaboration for creative work, in which humans and systems collaborate to create new value, will be essential for future dialogue systems. In this study, we collected 500 dialogues of human-human collaboration in Minecraft as a basis for developing a dialogue system that enables creative collaborative work. We conceived the Collaborative Garden Task, where two workers interact and collaborate in Minecraft to create a garden, and we collected dialogue, action logs, and subjective evaluations. We also collected third-person evaluations of the gardens and analyzed the relationship between dialogue and collaborative work that received high scores on the subjective and third-person evaluations in order to identify dialogic factors for high-quality collaborative work. We found that two essential aspects in creative collaborative work are performing more processes to ask for and agree on suggestions between workers and agreeing on a particular image of the final product in the early phase of work and then discussing changes and details.
Despite recent advances, dialogue systems still struggle to achieve fully autonomous transactions. Therefore, when a system encounters a problem, human operators need to take over the dialogue to complete the transaction. However, it is unclear what information should be presented to the operator when this handover takes place. In this study, we conducted a data collection experiment in which one of two operators talked to a user and switched with the other operator periodically while exchanging notes when the handovers took place. By examining these notes, it is possible to identify the information necessary for handing over the dialogue. We collected 60 dialogues in which two operators switched periodically while performing chat, consultation, and sales tasks in dialogue. We found that adjacency pairs are a useful representation for recording conversation history. In addition, we found that key-value-pair representation is also useful when there are underlying tasks, such as consultation and sales.
In this paper, we present the methodology of corpus design that will be used to study the comparison of influence between linguistic nudges with positive or negative influences and three conversational agents: robot, smart speaker, and human. We recruited forty-nine participants to form six groups. The conversational agents first asked the participants about their willingness to adopt five ecological habits and invest time and money in ecological problems. The participants were then asked the same questions but preceded by one linguistic nudge with positive or negative influence. The comparison of standard deviation and mean metrics of differences between these two notes (before the nudge and after) showed that participants were mainly affected by nudges with positive influence, even though several nudges with negative influence decreased the average note. In addition, participants from all groups were willing to spend more money than time on ecological problems. In general, our experiment’s early results suggest that a machine agent can influence participants to the same degree as a human agent. A better understanding of the power of influence of different conversational machines and the potential of influence of nudges of different polarities will lead to the development of ethical norms of human-computer interactions.
Building common ground with users is essential for dialogue agent systems and robots to interact naturally with people. While a few previous studies have investigated the process of building common ground in human-human dialogue, most of them have been conducted on the basis of text chat. In this study, we constructed a dialogue corpus to investigate the process of building common ground with a particular focus on the modality of dialogue and the social relationship between the participants in the process of building common ground, which are important but have not been investigated in the previous work. The results of our analysis suggest that adding the modality or developing the relationship between workers speeds up the building of common ground. Specifically, regarding the modality, the presence of video rather than only audio may unconsciously facilitate work, and as for the relationship, it is easier to convey information about emotions and turn-taking among friends than in first meetings. These findings and the corpus should prove useful for developing a system to support remote communication.
The ability to recognise emotions lends a conversational artificial intelligence a human touch. While emotions in chit-chat dialogues have received substantial attention, emotions in task-oriented dialogues remain largely unaddressed. This is despite emotions and dialogue success having equally important roles in a natural system. Existing emotion-annotated task-oriented corpora are limited in size, label richness, and public availability, creating a bottleneck for downstream tasks. To lay a foundation for studies on emotions in task-oriented dialogues, we introduce EmoWOZ, a large-scale manually emotion-annotated corpus of task-oriented dialogues. EmoWOZ is based on MultiWOZ, a multi-domain task-oriented dialogue dataset. It contains more than 11K dialogues with more than 83K emotion annotations of user utterances. In addition to Wizard-of-Oz dialogues from MultiWOZ, we collect human-machine dialogues within the same set of domains to sufficiently cover the space of various emotions that can happen during the lifetime of a data-driven dialogue system. To the best of our knowledge, this is the first large-scale open-source corpus of its kind. We propose a novel emotion labelling scheme, which is tailored to task-oriented dialogues. We report a set of experimental results to show the usability of this corpus for emotion recognition and state tracking in task-oriented dialogues.
Contextually aware intelligent agents are often required to understand the users and their surroundings in real-time. Our goal is to build Artificial Intelligence (AI) systems that can assist children in their learning process. Within such complex frameworks, Spoken Dialogue Systems (SDS) are crucial building blocks to handle efficient task-oriented communication with children in game-based learning settings. We are working towards a multimodal dialogue system for younger kids learning basic math concepts. Our focus is on improving the Natural Language Understanding (NLU) module of the task-oriented SDS pipeline with limited datasets. This work explores the potential benefits of data augmentation with paraphrase generation for the NLU models trained on small task-specific datasets. We also investigate the effects of extracting entities for conceivably further data expansion. We have shown that paraphrasing with model-in-the-loop (MITL) strategies using small seed data is a promising approach yielding improved performance results for the Intent Recognition task.
To build a well-founded opinion it is natural for humans to gather and exchange new arguments. Especially when being confronted with an overwhelming amount of information, people tend to focus on only the part of the available information that fits into their current beliefs or convenient opinions. To overcome this “self-imposed filter bubble” (SFB) in the information seeking process, it is crucial to identify influential indicators for the former. Within this paper we propose and investigate indicators for the the user’s SFB, mainly their Reflective User Engagement (RUE), their Personal Relevance (PR) ranking of content-related subtopics as well as their False (FK) and True Knowledge (TK) on the topic. Therefore, we analysed the answers of 202 participants of an online conducted user study, who interacted with our argumentative dialogue system BEA (“Building Engaging Argumentation”). Moreover, also the influence of different input/output modalities (speech/speech and drop-down menu/text) on the interaction with regard to the suggested indicators was investigated.
The COVID-19 pandemic and other global health events are unfortunately excellent environments for the creation and spread of misinformation, and the language associated with health misinformation may be typified by unique patterns and linguistic markers. Allowing health misinformation to spread unchecked can have devastating ripple effects; however, detecting and stopping its spread requires careful analysis of these linguistic characteristics at scale. We analyze prior investigations focusing on health misinformation, associated datasets, and detection of misinformation during health crises. We also introduce a novel dataset designed for analyzing such phenomena, comprised of 2.8 million news articles and social media posts spanning the early 1900s to the present. Our annotation guidelines result in strong agreement between independent annotators. We describe our methods for collecting this data and follow this with a thorough analysis of the themes and linguistic features that appear in information versus misinformation. Finally, we demonstrate a proof-of-concept misinformation detection task to establish dataset validity, achieving a strong performance benchmark (accuracy = 75%; F1 = 0.7).
We target the complementary binary tasks of identifying whether a tweet is misogynous and, if that is the case, whether it is also aggressive. We compare two ways to address these problems: one multi-class model that discriminates between all the classes at once: not misogynous, non aggressive-misogynous and aggressive-misogynous; as well as a cascaded approach where the binary classification is carried out separately (misog