Nous présentons un résumé en français et un résumé en anglais de l’article Is Attention Explanation ? An Introduction to the Debate (Bibal et al., 2022), publié dans les actes de la conférence 60th Annual Meeting of the Association for Computational Linguistics (ACL 2022).

The performance of deep learning models in NLP and other fields of machine learning has led to a rise in their popularity, and so the need for explanations of these models becomes paramount. Attention has been seen as a solution to increase performance, while providing some explanations. However, a debate has started to cast doubt on the explanatory power of attention in neural networks. Although the debate has created a vast literature thanks to contributions from various areas, the lack of communication is becoming more and more tangible. In this paper, we provide a clear overview of the insights on the debate by critically confronting works from these different areas. This holistic vision can be of great interest for future works in all the communities concerned by this debate. We sum up the main challenges spotted in these areas, and we conclude by discussing the most promising future avenues on attention as an explanation.

Evaluating automatic text simplification (ATS) systems is a difficult task that is either performed by automatic metrics or user-based evaluations. However, from a linguistic point-of-view, it is not always clear on what bases these evaluations operate. In this paper, we propose annotations of the ASSET corpus that can be used to shed more light on ATS evaluation. In addition to contributing with this resource, we show how it can be used to analyze SARI’s behavior and to re-evaluate existing ATS systems. We present our insights as a step to improve ATS evaluation protocols in the future.

In this article, we present an exploratory study on perceived word sense difficulty by native and non-native speakers of French. We use a graded lexicon in conjunction with the French Wiktionary to generate tasks in bundles of four items. Annotators manually rate the difficulty of the word senses based on their usage in a sentence by selecting the easiest and the most difficult word sense out of four. Our results show that the native and non-native speakers largely agree when it comes to the difficulty of words. Further, the rankings derived from the manual annotation broadly follow the levels of the words in the graded resource, although these levels were not overtly available to annotators. Using clustering, we investigate whether there is a link between the complexity of a definition and the difficulty of the associated word sense. However, results were inconclusive. The annotated data set is available for research purposes.

Lexical simplification is the task of substituting a difficult word with a simpler equivalent for a target audience. This is currently commonly done by modeling lexical complexity on a continuous scale to identify simpler alternatives to difficult words. In the TSAR shared task, the organizers call for systems capable of generating substitutions in a zero-shot-task context, for English, Spanish and Portuguese. In this paper, we present the solution we (the {textsc{cental} team) proposed for the task. We explore the ability of BERT-like models to generate substitution words by masking the difficult word. To do so, we investigate various context enhancement strategies, that we combined into an ensemble method. We also explore different substitution ranking methods. We report on a post-submission analysis of the results and present our insights for potential improvements. The code for all our experiments is available at https://gitlab.com/Cental-FR/cental-tsar2022.

Nous présentons un résumé en français et un résumé en anglais de l’article (Cardon & Grabar, 2020), publié dans les actes de la conférence 28th International Conference on Computational Linguistics (COLING 2020).

This paper describes and evaluates simple techniques for reducing the research space for parallel sentences in monolingual comparable corpora. Initially, when searching for parallel sentences between two comparable documents, all the possible sentence pairs between the documents have to be considered, which introduces a great degree of imbalance between parallel pairs and non-parallel pairs. This is a problem because even with a high performing algorithm, a lot of noise will be present in the extracted results, thus introducing a need for an extensive and costly manual check phase. We work on a manually annotated subset obtained from a French comparable corpus and show how we can drastically reduce the number of sentence pairs that have to be fed to a classifier so that the results can be manually handled.

L’édition 2020 du défi fouille de texte (DEFT) a proposé deux tâches autour de la similarité textuelle et une tâche d’extraction d’information. La première tâche vise à identifier le degré de similarité entre paires de phrases sur une échelle de 0 (le moins similaire) à 5 (le plus similaire). Les résultats varient de 0,65 à 0,82 d’EDRM. La deuxième tâche consiste à déterminer la phrase la plus proche d’une phrase source parmi trois phrases cibles fournies, avec des résultats très élevés, variant de 0,94 à 0,99 de précision. Ces deux tâches reposent sur un corpus du domaine général et de santé. La troisième tâche propose d’extraire dix catégories d’informations du domaine médical depuis le corpus de cas cliniques de DEFT 2019. Les résultats varient de 0,07 à 0,66 de F-mesure globale pour la sous-tâche des pathologies et signes ou symptômes, et de 0,14 à 0,76 pour la sous-tâche sur huit catégories médicales. Les méthodes utilisées reposent sur des CRF et des réseaux de neurones.

Semantic similarity is an area of Natural Language Processing that is useful for several downstream applications, such as machine translation, natural language generation, information retrieval, or question answering. The task consists in assessing the extent to which two sentences express or do not express the same meaning. To do so, corpora with graded pairs of sentences are required. The grade is positioned on a given scale, usually going from 0 (completely unrelated) to 5 (equivalent semantics). In this work, we introduce such a corpus for French, the first that we know of. It is comprised of 1,010 sentence pairs with grades from five annotators. We describe the annotation process, analyse these data, and perform a few experiments for the automatic grading of semantic similarity.

We present experiments on biomedical text simplification in French. We use two kinds of corpora – parallel sentences extracted from existing health comparable corpora in French and WikiLarge corpus translated from English to French – and a lexicon that associates medical terms with paraphrases. Then, we train neural models on these parallel corpora using different ratios of general and specialized sentences. We evaluate the results with BLEU, SARI and Kandel scores. The results point out that little specialized data helps significantly the simplification.

Parallel sentences provide semantically similar information which can vary on a given dimension, such as language or register. Parallel sentences with register variation (like expert and non-expert documents) can be exploited for the automatic text simplification. The aim of automatic text simplification is to better access and understand a given information. In the biomedical field, simplification may permit patients to understand medical and health texts. Yet, there is currently no such available resources. We propose to exploit comparable corpora which are distinguished by their registers (specialized and simplified versions) to detect and align parallel sentences. These corpora are in French and are related to the biomedical area. Manually created reference data show 0.76 inter-annotator agreement. Our purpose is to state whether a given pair of specialized and simplified sentences is parallel and can be aligned or not. We treat this task as binary classification (alignment/non-alignment). We perform experiments with a controlled ratio of imbalance and on the highly unbalanced real data. Our results show that the method we present here can be used to automatically generate a corpus of parallel sentences from our comparable corpus.

Les phrases parallèles contiennent des informations identiques ou très proches sémantiquement et offrent des indications importantes sur le fonctionnement de la langue. Lorsque les phrases sont différenciées par leur registre (comme expert vs. non-expert), elles peuvent être exploitées pour la simplification automatique de textes. Le but de la simplification automatique est d’améliorer la compréhension de textes. Par exemple, dans le domaine biomédical, la simplification peut permettre aux patients de mieux comprendre les textes relatifs à leur santé. Il existe cependant très peu de ressources pour la simplification en français. Nous proposons donc d’exploiter des corpus comparables, différenciés par leur technicité, pour y détecter des phrases parallèles et les aligner. Les données de référence sont créées manuellement et montrent un accord inter-annotateur de 0,76. Nous expérimentons sur des données équilibrées et déséquilibrées. La F-mesure sur les données équilibrées atteint jusqu’à 0,94. Sur les données déséquilibrées, les résultats sont plus faibles (jusqu’à 0,92 de F-mesure) mais restent compétitifs lorsque les modèles sont entraînés sur les données équilibrées.

The purpose of automatic text simplification is to transform technical or difficult to understand texts into a more friendly version. The semantics must be preserved during this transformation. Automatic text simplification can be done at different levels (lexical, syntactic, semantic, stylistic...) and relies on the corresponding knowledge and resources (lexicon, rules...). Our objective is to propose methods and material for the creation of transformation rules from a small set of parallel sentences differentiated by their technicity. We also propose a typology of transformations and quantify them. We work with French-language data related to the medical domain, although we assume that the method can be exploited on texts in any language and from any domain.

Notre travail traite de la simplification automatique de textes. Ce type d’application vise à rendre des contenus difficiles à comprendre plus lisibles. À partir de trois corpus comparables du domaine médical, d’un lexique existant et d’une terminologie du domaine, nous procédons à des analyses et à des modifications en vue de la simplification lexicale de textes médicaux. L’alignement manuel des phrases provenant de ces corpus comparables fournit des données de référence et permet d’analyser les procédés de simplification mis en place. La substitution lexicale avec la ressource existante permet d’effectuer de premiers tests de simplification lexicale et indique que des ressources plus spécifiques sont nécessaires pour traiter les textes médicaux. L’évaluation des substitutions est effectuée avec trois critères : grammaticalité, simplification et sémantique. Elle indique que la grammaticalité est plutôt bien sauvegardée, alors que la sémantique et la simplicité sont plus difficiles à gérer lors des substitutions avec ce type de méthodes.

Parallel aligned sentences provide useful information for different NLP applications. Yet, this kind of data is seldom available, especially for languages other than English. We propose to exploit comparable corpora in French which are distinguished by their registers (specialized and simplified versions) to detect and align parallel sentences. These corpora are related to the biomedical area. Our purpose is to state whether a given pair of specialized and simplified sentences is to be aligned or not. Manually created reference data show 0.76 inter-annotator agreement. We exploit a set of features and several automatic classifiers. The automatic alignment reaches up to 0.93 Precision, Recall and F-measure. In order to better evaluate the method, it is applied to data in English from the SemEval STS competitions. The same features and models are applied in monolingual and cross-lingual contexts, in which they show up to 0.90 and 0.73 F-measure, respectively.