Language models encode extensive factual knowledge within their parameters. The accurate assessment of this knowledge is crucial for understanding and improving these models. In the literature, factual knowledge assessment often relies on cloze sentences, which can lead to erroneous conclusions due to the complexity of natural language (out-of-subject continuations, the existence of many correct answers and the several ways of expressing them). In this paper, we introduce a new interpretable knowledge assessment method that mitigates these issues by leveraging distractors—incorrect but plausible alternatives to the correct answer. We propose several strategies for retrieving distractors and determine the most effective one through experimentation. Our method is evaluated against existing approaches, demonstrating solid alignment with human judgment and stronger robustness to verbalization artifacts. The code and data to reproduce our experiments are available on GitHub.

The scaling of causal language models in size and training data enabled them to tackle increasingly complex tasks. Despite the development of sophisticated tests to reveal their new capabilities, the underlying basis of these complex skills remains unclear. We argue that complex skills might be explained using simpler ones, represented by linguistic concepts. As an initial step in exploring this hypothesis, we focus on the lexical-semantic concept of synonymy, laying the groundwork for research into its relationship with more complex skills. We develop a comprehensive test suite to assess various aspects of synonymy under different conditions, and evaluate causal open-source models ranging up to 10 billion parameters. We find that these models effectively recognize synonymy but struggle to generate synonyms when prompted with relevant context.

Ce papier explore la robustesse des modèles de langue (ML) face aux variations du contexte temporel dans les connaissances factuelles. Il examine si les ML peuvent associer correctement un contexte temporel à un fait passé valide sur une période de temps délimitée, en leur demandant de différencier les contextes corrects des contextes incorrects. La capacité de distinction des ML est analysée sur deux dimensions : la distance du contexte incorrect par rapport à la période de validité et la granularité du contexte. Pour cela, un jeu de données, TimeStress, est introduit, permettant de tester 18 ML variés. Les résultats révèlent que le meilleur ML n’atteint une distinction parfaite que pour 11% des faits étudiés, avec des erreurs critiques qu’un humain ne ferait pas. Ces travaux soulignent les limites des ML actuels en matière de représentation temporelle.

This paper explores the robustness of language models (LMs) to variations in the temporal context within factual knowledge. It examines whether LMs can correctly associate a temporal context with a past fact valid over a defined period, by asking them to differentiate correct from incorrect contexts. The LMs’ ability to distinguish is analyzed along two dimensions: the distance of the incorrect context from the validity period and the granularity of the context. To this end, a dataset called TimeStress is introduced, enabling the evaluation of 18 diverse LMs. Results reveal that the best LM achieves a perfect distinction for only 11% of the studied facts, with errors, certainly rare, but critical that humans would not make. This work highlights the limitations of current LMs in temporal representation.

La factualité des modèles de langue se dégrade avec le temps puisque les événements postérieurs à leur entraînement leur sont inconnus. Une façon de maintenir ces modèles à jour pourrait être la mise à jour factuelle à l’échelle de faits atomiques. Pour étudier cette tâche, nous présentons WikiFactDiff, un jeu de données qui représente les changements survenus entre deux dates sous la forme d’un ensemble de faits simples, sous format RDF, divisés en trois catégories : les faits à apprendre, les faits à conserver et les faits obsolètes. Ces faits sont verbalisés afin de permettre l’exécution des algorithmes de mise à jour et leur évaluation, qui est présentée dans ce document. Contrairement aux jeux de données existants, WikiFactDiff représente un cadre de mise à jour réaliste qui implique divers scénarios, notamment les remplacements de faits, leur archivage et l’insertion de nouvelles entités.

The factuality of large language model (LLMs) tends to decay over time since events posterior to their training are “unknown” to them. One way to keep models up-to-date could be factual update: the task of inserting, replacing, or removing certain simple (atomic) facts within the model. To study this task, we present WikiFactDiff, a dataset that describes the evolution of factual knowledge between two dates as a collection of simple facts divided into three categories: new, obsolete, and static. We describe several update scenarios arising from various combinations of these three types of basic update. The facts are represented by subject-relation-object triples; indeed, WikiFactDiff was constructed by comparing the state of the Wikidata knowledge base at 4 January 2021 and 27 February 2023. Those fact are accompanied by verbalization templates and cloze tests that enable running update algorithms and their evaluation metrics. Contrary to other datasets, such as zsRE and CounterFact, WikiFactDiff constitutes a realistic update setting that involves various update scenarios, including replacements, archival, and new entity insertions. We also present an evaluation of existing update algorithms on WikiFactDiff.

Text generation from Abstract Meaning Representation (AMR) has substantially benefited from the popularized Pretrained Language Models (PLMs). Myriad approaches have linearized the input graph as a sequence of tokens to fit the PLM tokenization requirements. Nevertheless, this transformation jeopardizes the structural integrity of the graph and is therefore detrimental to its resulting representation. To overcome this issue, Ribeiro et al. (2021b) have recently proposed StructAdapt, a structure-aware adapter which injects the input graph connectivity within PLMs using Graph Neural Networks (GNNs). In this paper, we investigate the influence of Relative Position Embeddings (RPE) on AMR-to-Text, and, in parallel, we examine the robustness of StructAdapt. Through ablation studies, graph attack and link prediction, we reveal that RPE might be partially encoding input graphs. We suggest further research regarding the role of RPE will provide valuable insights for Graph-to-Text generation.

Structured Knowledge has recently emerged as an essential component to support fine-grained Question Answering (QA). In general, QA systems query a Knowledge Base (KB) to detect and extract the raw answers as final prediction. However, as lacking of context, language generation can offer a much informative and complete response. In this paper, we propose to combine the power of transfer learning and the advantage of entity placeholders to produce high-quality verbalization of extracted answers from a KB. We claim that such approach is especially well-suited for answer generation. Our experiments show 44.25%, 3.26% and 29.10% relative gain in BLEU over the state-of-the-art on the VQuAnDA, ParaQA and VANiLLa datasets, respectively. We additionally provide minor hallucinations corrections in VANiLLa standing for 5% of each of the training and testing set. We witness a median absolute gain of 0.81 SacreBLEU. This strengthens the importance of data quality when using automated evaluation.

Greedy algorithms for NLP such as transition-based parsing are prone to error propagation. One way to overcome this problem is to allow the algorithm to backtrack and explore an alternative solution in cases where new evidence contradicts the solution explored so far. In order to implement such a behavior, we use reinforcement learning and let the algorithm backtrack in cases where such an action gets a better reward than continuing to explore the current solution. We test this idea on both POS tagging and dependency parsing and show that backtracking is an effective means to fight against error propagation.

In this paper we describe our contribution to the CMCL 2021 Shared Task, which consists in predicting 5 different eye tracking variables from English tokenized text. Our approach is based on a neural network that combines both raw textual features we extracted from the text and parser-based features that include linguistic predictions (e.g. part of speech) and complexity metrics (e.g., entropy of parsing). We found that both the features we considered as well as the architecture of the neural model that combined these features played a role in the overall performance. Our system achieved relatively high accuracy on the test data of the challenge and was ranked 2nd out of 13 competing teams and a total of 30 submissions.

The Reading Machine, is a parsing framework that takes as input raw text and performs six standard nlp tasks: tokenization, pos tagging, morphological analysis, lemmatization, dependency parsing and sentence segmentation. It is built upon Transition Based Parsing, and allows to implement a large number of parsing configurations, among which a fully incremental one. Three case studies are presented to highlight the versatility of the framework. The first one explores whether an incremental parser is able to take into account top-down dependencies (i.e. the influence of high level decisions on low level ones), the second compares the performances of an incremental and a pipe-line architecture and the third quantifies the impact of the right context on the predictions made by an incremental parser.

Contextualised embeddings such as BERT have become de facto state-of-the-art references in many NLP applications, thanks to their impressive performances. However, their opaqueness makes it hard to interpret their behaviour. SLICE is a hybrid model that combines supersense labels with contextual embeddings. We introduce a weakly supervised method to learn interpretable embeddings from raw corpora and small lists of seed words. Our model is able to represent both a word and its context as embeddings into the same compact space, whose dimensions correspond to interpretable supersenses. We assess the model in a task of supersense tagging for French nouns. The little amount of supervision required makes it particularly well suited for low-resourced scenarios. Thanks to its interpretability, we perform linguistic analyses about the predicted supersenses in terms of input word and context representations.

La compréhension automatique de texte est une tâche faisant partie de la famille des systèmes de Question/Réponse où les questions ne sont pas à portée générale mais sont liées à un document particulier. Récemment de très grand corpus (SQuAD, MS MARCO) contenant des triplets (document, question, réponse) ont été mis à la disposition de la communauté scientifique afin de développer des méthodes supervisées à base de réseaux de neurones profonds en obtenant des résultats prometteurs. Ces méthodes sont cependant très gourmandes en données d’apprentissage, données qui n’existent pour le moment que pour la langue anglaise. Le but de cette étude est de permettre le développement de telles ressources pour d’autres langues à moindre coût en proposant une méthode générant de manière semi-automatique des questions à partir d’une analyse sémantique d’un grand corpus. La collecte de questions naturelle est réduite à un ensemble de validation/test. L’application de cette méthode sur le corpus CALOR-Frame a permis de développer la ressource CALOR-QUEST présentée dans cet article.

Machine reading comprehension is a task related to Question-Answering where questions are not generic in scope but are related to a particular document. Recently very large corpora (SQuAD, MS MARCO) containing triplets (document, question, answer) were made available to the scientific community to develop supervised methods based on deep neural networks with promising results. These methods need very large training corpus to be efficient, however such kind of data only exists for English and Chinese at the moment. The aim of this study is the development of such resources for other languages by proposing to generate in a semi-automatic way questions from the semantic Frame analysis of large corpora. The collect of natural questions is reduced to a validation/test set. We applied this method on the CALOR-Frame French corpus to develop the CALOR-QUEST resource presented in this paper.

The existence of universal models to describe the syntax of languages has been debated for decades. The availability of resources such as the Universal Dependencies treebanks and the World Atlas of Language Structures make it possible to study the plausibility of universal grammar from the perspective of dependency parsing. Our work investigates the use of high-level language descriptions in the form of typological features for multilingual dependency parsing. Our experiments on multilingual parsing for 40 languages show that typological information can indeed guide parsers to share information between similar languages beyond simple language identification.

La désambiguïsation des rattachements prépositionnels est une tâche syntaxique qui demande des connaissances sémantiques, pouvant être extraites d’une image associée au texte traité. Nous présentons et analysons les difficultés de cette tâche pour laquelle nous construisons un système complet entraîné sur une version étendue des annotations du corpus Flickr30k Entities. Lorsque la sémantique lexicale n’est pas disponible, l’information visuelle apporte 3 % d’amélioration.

Les conversations techniques en ligne sont un type de productions linguistiques qui par de nombreux aspects se démarquent des objets plus usuellement étudiés en traitement automatique des langues : il s’agit de dialogues écrits entre deux locuteurs qui servent de support à la résolution coopérative des problèmes des usagers. Nous proposons de décrire ici ces conversations par un étiquetage en actes de dialogue spécifiquement conçu pour les conversations en ligne. Différents systèmes de prédictions ont été évalués ainsi qu’une méthode permettant de s’abstraire des spécificités lexicales du corpus d’apprentissage.

We present supertagging-based models for Tree Adjoining Grammar parsing that use neural network architectures and dense vector representation of supertags (elementary trees) to achieve state-of-the-art performance in unlabeled and labeled attachment scores. The shift-reduce parsing model eschews lexical information entirely, and uses only the 1-best supertags to parse a sentence, providing further support for the claim that supertagging is “almost parsing.” We demonstrate that the embedding vector representations the parser induces for supertags possess linguistically interpretable structure, supporting analogies between grammatical structures like those familiar from recent work in distributional semantics. This dense representation of supertags overcomes the drawbacks for statistical models of TAG as compared to CCG parsing, raising the possibility that TAG is a viable alternative for NLP tasks that require the assignment of richer structural descriptions to sentences.

PP-attachments are an important source of errors in parsing natural language. We propose in this article to use data coming from a multimodal corpus, combining textual, visual and conceptual information, as well as a correction strategy, to propose alternative attachments in the output of a parser.

Syntax plays an important role in the task of predicting the semantic structure of a sentence. But syntactic phenomena such as alternations, control and raising tend to obfuscate the relation between syntax and semantics. In this paper we predict the semantic structure of a sentence using a deeper syntax than what is usually done. This deep syntactic representation abstracts away from purely syntactic phenomena and proposes a structural organization of the sentence that is closer to the semantic representation. Experiments conducted on a French corpus annotated with semantic frames showed that a semantic parser reaches better performances with such a deep syntactic input.

We introduce DeQue, a lexicon covering French complex prepositions (CPRE) like “à partir de” (from) and complex conjunctions (CCONJ) like “bien que” (although). The lexicon includes fine-grained linguistic description based on empirical evidence. We describe the general characteristics of CPRE and CCONJ in French, with special focus on syntactic ambiguity. Then, we list the selection criteria used to build the lexicon and the corpus-based methodology employed to collect entries. Finally, we quantify the ambiguity of each construction by annotating around 100 sentences randomly taken from the FRWaC. In addition to its theoretical value, the resource has many potential practical applications. We intend to employ DeQue for treebank annotation and to train a dependency parser that can takes complex constructions into account.

Syntactic parsing of speech transcriptions faces the problem of the presence of disfluencies that break the syntactic structure of the utterances. We propose in this paper two solutions to this problem. The first one relies on a disfluencies predictor that detects disfluencies and removes them prior to parsing. The second one integrates the disfluencies in the syntactic structure of the utterances and train a disfluencies aware parser.

This paper describes the syntactic annotation process of the DECODA corpus. This corpus contains manual transcriptions of spoken conversations recorded in the French call-center of the Paris Public Transport Authority (RATP). Three levels of syntactic annotation have been performed with a semi-supervised approach: POS tags, Syntactic Chunks and Dependency parses. The main idea is to use off-the-shelf NLP tools and models, originaly developped and trained on written text, to perform a first automatic annotation on the manually transcribed corpus. At the same time a fully manual annotation process is performed on a subset of the original corpus, called the GOLD corpus. An iterative process is then applied, consisting in manually correcting errors found in the automatic annotations, retraining the linguistic models of the NLP tools on this corrected corpus, then checking the quality of the adapted models on the fully manual annotations of the GOLD corpus. This process iterates until a certain error rate is reached. This paper describes this process, the main issues raising when adapting NLP tools to process speech transcriptions, and presents the first evaluations performed with these new adapted tools.

La fouille de données orales est un domaine de recherche visant à caractériser un flux audio contenant de la parole d’un ou plusieurs locuteurs, à l’aide de descripteurs liés à la forme et au contenu du signal. Outre la transcription automatique en mots des paroles prononcées, des informations sur le type de flux audio traité ainsi que sur le rôle et l’identité des locuteurs sont également cruciales pour permettre des requêtes complexes telles que : « chercher des débats sur le thème X », « trouver toutes les interviews de Y », etc. Dans ce cadre, et en traitant des conversations enregistrées lors d’émissions de radio ou de télévision, nous étudions la manière dont les locuteurs expriment des questions dans les conversations, en partant de l’intuition initiale que la forme des questions posées est une signature du rôle du locuteur dans la conversation (présentateur, invité, auditeur, etc.). En proposant une classification du type des questions et en utilisant ces informations en complément des descripteurs généralement utilisés dans la littérature pour classer les locuteurs par rôle, nous espérons améliorer l’étape de classification, et valider par la même occasion notre intuition initiale.

Nous présentons une architecture pour l’analyse syntaxique en deux étapes. Dans un premier temps un analyseur syntagmatique construit, pour chaque phrase, une liste d’analyses qui sont converties en arbres de dépendances. Ces arbres sont ensuite réévalués par un réordonnanceur discriminant. Cette méthode permet de prendre en compte des informations auxquelles l’analyseur n’a pas accès, en particulier des annotations fonctionnelles. Nous validons notre approche par une évaluation sur le corpus arboré de Paris 7. La seconde étape permet d’améliorer significativement la qualité des analyses retournées, quelle que soit la métrique utilisée.

La constitution de ressources linguistiques est une tâche longue et coûteuse. C’est notamment le cas pour les ressources morphologiques. Ces ressources décrivent de façon approfondie et explicite l’organisation morphologique du lexique complétée d’informations sémantiques exploitables dans le domaine du TAL. Le travail que nous présentons dans cet article s’inscrit dans cette perspective et, plus particulièrement, dans l’optique d’affiner une ressource existante en s’appuyant sur des informations sémantiques obtenues automatiquement. Notre objectif est de caractériser sémantiquement des familles morpho-phonologiques (des mots partageant une même racine et une continuité de sens). Pour ce faire, nous avons utilisé des informations extraites du TLFi annoté morpho-syntaxiquement. Les premiers résultats de ce travail seront analysés et discutés.

Cet article décrit un modèle d’analyse syntaxique de l’oral spontané axé sur la reconnaissance de cadres valenciels verbaux. Le modèle d’analyse se décompose en deux étapes : une étape générique, basée sur des ressources génériques du français et une étape de réordonnancement des solutions de l’analyseur réalisé par un modèle spécifique à une application. Le modèle est évalué sur le corpus MEDIA.

Cet article présente une manière d’intégrer un étiqueteur morpho-syntaxique et un analyseur partiel. Cette integration permet de corriger des erreurs effectuées par l’étiqueteur seul. L’étiqueteur et l’analyseur ont été réalisés sous la forme d’automates pondérés. Des résultats sur un corpus du français ont montré une dimintion du taux d’erreur de l’ordre de 12%.

This paper reports on an experiment in assembling a domain-specific machine translation prototype system from off-the-shelf components. The design goals of this experiment were to reuse existing components, to use machine-learning techniques for parser specialization and for transfer lexicon extraction, and to use an expressive, lexicalized formalism for the transfer component.