This is an internal, incomplete preview of a proposed change to the ACL Anthology.
For efficiency reasons, we generate only three BibTeX files per volume, and the preview may be incomplete in other ways, or contain mistakes.
Do not treat this content as an official publication.
Linguistic data, a component critical not only for research in a variety of fields but also for the development of various Natural Language Processing (NLP) applications, can contain personal information. As a result, its accessibility is limited, both from a legal and an ethical standpoint. One of the solutions is the pseudonymization of the data. Key stages of this process include the identification of sensitive elements and the generation of suitable surrogates in a way that the data is still useful for the intended task. Within this paper, we conduct an analysis of tagsets that have previously been utilized in anonymization and pseudonymization. We also investigate what kinds of Personally Identifiable Information (PII) appear in various domains. These reveal that none of the analyzed tagsets account for all of the PII types present cross-domain at the level of detailedness seemingly required for pseudonymization. We advocate for a universal system of tags for categorizing PIIs leading up to their replacement. Such categorization could facilitate the generation of grammatically, semantically, and sociolinguistically appropriate surrogates for the kinds of information that are considered sensitive in a given domain, resulting in a system that would enable dynamic pseudonymization while keeping the texts readable and useful for future research in various fields.
Linguistic data can — and often does — contain PII (Personal Identifiable Information). Both from a legal and ethical standpoint, the sharing of such data is not permissible. According to the GDPR, pseudonymization, i.e. the replacement of sensitive information with surrogates, is an acceptable strategy for privacy preservation. While research has been conducted on the detection and replacement of sensitive data in Swedish medical data using Large Language Models (LLMs), it is unclear whether these models handle PII in less structured and more thematically varied texts equally well. In this paper, we present and discuss the performance of an LLM-based PII-detection system for Swedish learner essays.
Automated essay scoring (AES) of second-language learner essays is a high-stakes task as it can affect the job and educational opportunities a student may have access to. Thus, it becomes imperative to make sure that the essays are graded based on the students’ language proficiency as opposed to other reasons, such as personal names used in the text of the essay. Moreover, most of the research data for AES tends to contain personal identifiable information. Because of that, pseudonymization becomes an important tool to make sure that this data can be freely shared. Thus, our systems should not grade students based on which given names were used in the text of the essay, both for fairness and for privacy reasons. In this paper we explore how given names affect the CEFR level classification of essays of second language learners of Swedish. We use essays containing just one personal name and substitute it for names from lists of given names from four different ethnic origins, namely Swedish, Finnish, Anglo-American, and Arabic. We find that changing the names within the essays has no apparent effect on the classification task, regardless of whether a feature-based or a transformer-based model is used.
Generative language models have been used to study a wide variety of phenomena in NLP. This allows us to better understand the linguistic capabilities of those models and to better analyse the texts that we are working with. However, these studies have mainly focused on text generated by L1 speakers of English. In this paper we study whether linguistic competence of L2 learners of Swedish (through their performance on essay tasks) correlates with the perplexity of a decoder-only model (GPT-SW3). We run two sets of experiments, doing both quantitative and qualitative analyses for each of them. In the first one, we analyse the perplexities of the essays and compare them with the CEFR level of the essays, both from an essay-wide level and from a token level. In our second experiment, we compare the perplexity of an L2 learner essay with a normalised version of it. We find that the perplexity of essays tends to be lower for higher CEFR levels and that normalised essays have a lower perplexity than the original versions. Moreover, we find that different factors can lead to spikes in perplexity, not all of them being related to L2 learner language.
Language-and-vision models have shown good performance in tasks such as image-caption matching and caption generation. However, it is challenging for such models to generate pragmatically correct captions, which adequately reflect what is happening in one image or several images. It is crucial to evaluate this behaviour to understand underlying reasons behind it. Here we explore to what extent contextual language-and-vision models are sensitive to different discourse, both textual and visual. In particular, we employ one of the multi-modal transformers (ViLBERT) and test if it can match descriptions and images, differentiating them from distractors of different degree of similarity that are sampled from different visual and textual contexts. We place our evaluation in the multi-sentence and multi-image setup, where images and sentences are expected to form a single narrative structure. We show that the model can distinguish different situations but it is not sensitive to differences within one narrative structure. We also show that performance depends on the task itself, for example, what modality remains unchanged in non-matching pairs or how similar non-matching pairs are to original pairs.
We evaluate the role of expert-based domain knowledge and resources in relation to training large language models by referring to our work on training and evaluating neural models, also in under-resourced scenarios which we believe also informs training models for “well-resourced” languages and domains. We argue that our community needs both large-scale datasets and small but high-quality data based on expert knowledge and that both activities should work hand-in-hand.
We propose the VDG Challenge: a shared task that addresses and benchmarks the task of utterance generation in collaborative visual dialogue. The task features two challenging datasets, an evaluation protocol and a tentative schedule. Our shared task will allow researchers to unravel problems of modelling multi-modal interaction and fit of the existing approaches in the NLP and NLG communities.
We present a first release of 500 documents from the multimodal corpus Tell-me-more (Ilinykh et al., 2019) annotated with coreference information according to the ARRAU guidelines (Poesio et al., 2021). The corpus consists of images and short texts of five sentences. We describe the annotation process and present the adaptations to the original guidelines in order to account for the challenges of grounding the annotations to the image. 50 documents from the 500 available are annotated by two people and used to estimate inter-annotator agreement (IAA) relying on Krippendorff’s alpha.
The usage of social media platforms has resulted in the proliferation of work on Arabic Natural Language Processing (ANLP), including the development of resources. There is also an increased interest in processing Arabic dialects and a number of models and algorithms have been utilised for the purpose of Dialectal Arabic Natural Language Processing (DANLP). In this paper, we conduct a comparison study between some of the most well-known and most commonly used methods in NLP in order to test their performance on different corpora and two NLP tasks: Dialect Identification and Sentiment Analysis. In particular, we compare three general classes of models: a) traditional Machine Learning models with features, b) classic Deep Learning architectures (LSTMs) with pre-trained word embeddings and lastly c) different Bidirectional Encoder Representations from Transformers (BERT) models such as (Multilingual-BERT, Ara-BERT, and Twitter-Arabic-BERT). The results of the comparison show that using feature-based classification can still compete with BERT models in these dialectal Arabic contexts. The use of transformer models have the ability to outperform traditional Machine Learning approaches, depending on the type of text they have been trained on, in contrast to classic Deep Learning models like LSTMs which do not perform well on the tasks
In this paper we examine different meaning representations that are commonly used in different natural language applications today and discuss their limits, both in terms of the aspects of the natural language meaning they are modelling and in terms of the aspects of the application for which they are used.
In this paper, we present a number of fine-grained resources for Natural Language Inference (NLI). In particular, we present a number of resources and validation methods for Greek NLI and a resource for precise NLI. First, we extend the Greek version of the FraCaS test suite to include examples where the inference is directly linked to the syntactic/morphological properties of Greek. The new resource contains an additional 428 examples, making it in total a dataset of 774 examples. Expert annotators have been used in order to create the additional resource, while extensive validation of the original Greek version of the FraCaS by non-expert and expert subjects is performed. Next, we continue the work initiated by (CITATION), according to which a subset of the RTE problems have been labeled for missing hypotheses and we present a dataset an order of magnitude larger, annotating the whole SuperGlUE/RTE dataset with missing hypotheses. Lastly, we provide a de-dropped version of the Greek XNLI dataset, where the pronouns that are missing due to the pro-drop nature of the language are inserted. We then run some models to see the effect of that insertion and report the results.
We investigate how different augmentation techniques on both textual and visual representations affect the performance of the face description generation model. Specifically, we provide the model with either original images, sketches of faces, facial composites or distorted images. In addition, on the language side, we experiment with different methods to augment the original dataset with paraphrased captions, which are semantically equivalent to the original ones, but differ in terms of their form. We also examine if augmenting the dataset with descriptions from a different domain (e.g., image captions of real-world images) has an effect on the performance of the models. We train models on different combinations of visual and linguistic features and perform both (i) automatic evaluation of generated captions and (ii) examination of how useful different visual features are for the task of facial feature classification. Our results show that although original images encode the best possible representation for the task, the model trained on sketches can still perform relatively well. We also observe that augmenting the dataset with descriptions from a different domain can boost performance of the model. We conclude that face description generation systems are more susceptible to language rather than vision data augmentation. Overall, we demonstrate that face caption generation models display a strong imbalance in the utilisation of language and vision modalities, indicating a lack of proper information fusion. We also describe ethical implications of our study and argue that future work on human face description generation should create better, more representative datasets.
This paper describes insights into how different inference algorithms structure discourse in image paragraphs. We train a multi-modal transformer and compare 11 variations of decoding algorithms. We propose to evaluate image paragraphs not only with standard automatic metrics, but also with a more extensive, “under the hood” analysis of the discourse formed by sentences. Our results show that while decoding algorithms can be unfaithful to the reference texts, they still generate grounded descriptions, but they also lack understanding of the discourse structure and differ from humans in terms of attentional structure over images.
We explore how a multi-modal transformer trained for generation of longer image descriptions learns syntactic and semantic representations about entities and relations grounded in objects at the level of masked self-attention (text generation) and cross-modal attention (information fusion). We observe that cross-attention learns the visual grounding of noun phrases into objects and high-level semantic information about spatial relations, while text-to-text attention captures low-level syntactic knowledge between words. This concludes that language models in a multi-modal task learn different semantic information about objects and relations cross-modally and uni-modally (text-only). Our code is available here: https://github.com/GU-CLASP/attention-as-grounding.
In recent years several corpora have been developed for vision and language tasks. We argue that there is still significant room for corpora that increase the complexity of both visual and linguistic domains and which capture different varieties of perceptual and conversational contexts. Working with two corpora approaching this goal, we present a linguistic perspective on some of the challenges in creating and extending resources combining language and vision while preserving continuity with the existing best practices in the area of coreference annotation.
The problem of interpretation of knowledge learned by multi-head self-attention in transformers has been one of the central questions in NLP. However, a lot of work mainly focused on models trained for uni-modal tasks, e.g. machine translation. In this paper, we examine masked self-attention in a multi-modal transformer trained for the task of image captioning. In particular, we test whether the multi-modality of the task objective affects the learned attention patterns. Our visualisations of masked self-attention demonstrate that (i) it can learn general linguistic knowledge of the textual input, and (ii) its attention patterns incorporate artefacts from visual modality even though it has never accessed it directly. We compare our transformer’s attention patterns with masked attention in distilgpt-2 tested for uni-modal text generation of image captions. Based on the maps of extracted attention weights, we argue that masked self-attention in image captioning transformer seems to be enhanced with semantic knowledge from images, exemplifying joint language-and-vision information in its attention patterns.
In recent years several corpora have been developed for vision and language tasks. With this paper, we intend to start a discussion on the annotation of referential phenomena in situated dialogue. We argue that there is still significant room for corpora that increase the complexity of both visual and linguistic domains and which capture different varieties of perceptual and conversational contexts. In addition, a rich annotation scheme covering a broad range of referential phenomena and compatible with the textual task of coreference resolution is necessary in order to take the most advantage of these corpora. Consequently, there are several open questions regarding the semantics of reference and annotation, and the extent to which standard textual coreference accounts for the situated dialogue genre. Working with two corpora on situated dialogue, we present our extension to the ARRAU (Uryupina et al., 2020) annotation scheme in order to start this discussion.
This work explores the differences and similarities between neural image classifiers’ mis-categorisations and visually grounded metaphors - that we could conceive as intentional mis-categorisations. We discuss the possibility of using automatic image classifiers to approximate human metaphoric behaviours, and the limitations of such frame. We report two pilot experiments to study grounded metaphoricity. In the first we represent metaphors as a form of visual mis-categorisation. In the second we model metaphors as a more flexible, compositional operation in a continuous visual space generated from automatic classification systems.
The major shortcomings of using neural networks with situated agents are that in incremental interaction very few learning examples are available and that their visual sensory representations are quite different from image caption datasets. In this work we adapt and evaluate a few-shot learning approach, Matching Networks (Vinyals et al., 2016), to conversational strategies of a robot interacting with a human tutor in order to efficiently learn to categorise objects that are presented to it and also investigate to what degree transfer learning from pre-trained models on images from different contexts can improve its performance. We discuss the implications of such learning on the nature of semantic representations the system has learned.
As the number of social media users increases, they express their thoughts, needs, socialise and publish their opinions reviews. For good social media sentiment analysis, good quality resources are needed, and the lack of these resources is particularly evident for languages other than English, in particular Arabic. The available Arabic resources lack of from either the size of the corpus or the quality of the annotation. In this paper, we present an Arabic Sentiment Analysis Corpus collected from Twitter, which contains 36K tweets labelled into positive and negative. We employed distant supervision and self-training approaches into the corpus to annotate it. Besides, we release an 8K tweets manually annotated as a gold standard. We evaluated the corpus intrinsically by comparing it to human classification and pre-trained sentiment analysis models, Moreover, we apply extrinsic evaluation methods exploiting sentiment analysis task and achieve an accuracy of 86%.
Generating multi-sentence image descriptions is a challenging task, which requires a good model to produce coherent and accurate paragraphs, describing salient objects in the image. We argue that multiple sources of information are beneficial when describing visual scenes with long sequences. These include (i) perceptual information and (ii) semantic (language) information about how to describe what is in the image. We also compare the effects of using two different pooling mechanisms on either a single modality or their combination. We demonstrate that the model which utilises both visual and language inputs can be used to generate accurate and diverse paragraphs when combined with a particular pooling mechanism. The results of our automatic and human evaluation show that learning to embed semantic information along with visual stimuli into the paragraph generation model is not trivial, raising a variety of proposals for future experiments.
Understanding and generating spatial descriptions requires knowledge about what objects are related, their functional interactions, and where the objects are geometrically located. Different spatial relations have different functional and geometric bias. The wide usage of neural language models in different areas including generation of image description motivates the study of what kind of knowledge is encoded in neural language models about individual spatial relations. With the premise that the functional bias of relations is expressed in their word distributions, we construct multi-word distributional vector representations and show that these representations perform well on intrinsic semantic reasoning tasks, thus confirming our premise. A comparison of our vector representations to human semantic judgments indicates that different bias (functional or geometric) is captured in different data collection tasks which suggests that the contribution of the two meaning modalities is dynamic, related to the context of the task.
We explore the extent to which neural networks can learn to identify semantically equivalent sentences from a small variable dataset using an end-to-end training. We collect a new noisy non-standardised user-generated Algerian (ALG) dataset and also translate it to Modern Standard Arabic (MSA) which serves as its regularised counterpart. We compare the performance of various models on both datasets and report the best performing configurations. The results show that relatively simple models composed of 2 LSTM layers outperform by far other more sophisticated attention-based architectures, for both ALG and MSA datasets.
Generating grounded image descriptions requires associating linguistic units with their corresponding visual clues. A common method is to train a decoder language model with attention mechanism over convolutional visual features. Attention weights align the stratified visual features arranged by their location with tokens, most commonly words, in the target description. However, words such as spatial relations (e.g. next to and under) are not directly referring to geometric arrangements of pixels but to complex geometric and conceptual representations. The aim of this paper is to evaluate what representations facilitate generating image descriptions with spatial relations and lead to better grounded language generation. In particular, we investigate the contribution of three different representational modalities in generating relational referring expressions: (i) pre-trained convolutional visual features, (ii) different top-down geometric relational knowledge between objects, and (iii) world knowledge captured by contextual embeddings in language models.
We work with Algerian, an under-resourced non-standardised Arabic variety, for which we compile a new parallel corpus consisting of user-generated textual data matched with normalised and corrected human annotations following data-driven and our linguistically motivated standard. We use an end-to-end deep neural model designed to deal with context-dependent spelling correction and normalisation. Results indicate that a model with two CNN sub-network encoders and an LSTM decoder performs the best, and that word context matters. Additionally, pre-processing data token-by-token with an edit-distance based aligner significantly improves the performance. We get promising results for the spelling correction and normalisation, as a pre-processing step for downstream tasks, on detecting binary Semantic Textual Similarity.
This paper seeks to examine the effect of including background knowledge in the form of character pre-trained neural language model (LM), and data bootstrapping to overcome the problem of unbalanced limited resources. As a test, we explore the task of language identification in mixed-language short non-edited texts with an under-resourced language, namely the case of Algerian Arabic for which both labelled and unlabelled data are limited. We compare the performance of two traditional machine learning methods and a deep neural networks (DNNs) model. The results show that overall DNNs perform better on labelled data for the majority categories and struggle with the minority ones. While the effect of the untokenised and unlabelled data encoded as LM differs for each category, bootstrapping, however, improves the performance of all systems and all categories. These methods are language independent and could be generalised to other under-resourced languages for which a small labelled data and a larger unlabelled data are available.
The challenge for computational models of spatial descriptions for situated dialogue systems is the integration of information from different modalities. The semantics of spatial descriptions are grounded in at least two sources of information: (i) a geometric representation of space and (ii) the functional interaction of related objects that. We train several neural language models on descriptions of scenes from a dataset of image captions and examine whether the functional or geometric bias of spatial descriptions reported in the literature is reflected in the estimated perplexity of these models. The results of these experiments have implications for the creation of models of spatial lexical semantics for human-robot dialogue systems. Furthermore, they also provide an insight into the kinds of the semantic knowledge captured by neural language models trained on spatial descriptions, which has implications for image captioning systems.
We explore the effect of injecting background knowledge to different deep neural network (DNN) configurations in order to mitigate the problem of the scarcity of annotated data when applying these models on datasets of low-resourced languages. The background knowledge is encoded in the form of lexicons and pre-trained sub-word embeddings. The DNN models are evaluated on the task of detecting code-switching and borrowing points in non-standardised user-generated Algerian texts. Overall results show that DNNs benefit from adding background knowledge. However, the gain varies between models and categories. The proposed DNN architectures are generic and could be applied to other low-resourced languages.
This paper presents a language identification system designed to detect the language of each word, in its context, in a multilingual documents as generated in social media by bilingual/multilingual communities, in our case speakers of Algerian Arabic. We frame the task as a sequence tagging problem and use supervised machine learning with standard methods like HMM and Ngram classification tagging. We also experiment with a lexicon-based method. Combining all the methods in a fall-back mechanism and introducing some linguistic rules, to deal with unseen tokens and ambiguous words, gives an overall accuracy of 93.14%. Finally, we introduced rules for language identification from sequences of recognised words.
Type theory has played an important role in specifying the formal connection between syntactic structure and semantic interpretation within the history of formal semantics. In recent years rich type theories developed for the semantics of programming languages have become influential in the semantics of natural language. The use of probabilistic reasoning to model human learning and cognition has become an increasingly important part of cognitive science. In this paper we offer a probabilistic formulation of a rich type theory, Type Theory with Records (TTR), and we illustrate how this framework can be used to approach the problem of semantic learning. Our probabilistic version of TTR is intended to provide an interface between the cognitive process of classifying situations according to the types that they instantiate, and the compositional semantics of natural language.