Pretrained language models (PLMs) display impressive performances and have captured the attention of the NLP community.Establishing best practices in pretraining has, therefore, become a major focus of NLP research, especially since insights gained from monolingual English models may not necessarily apply to more complex multilingual models.One significant caveat of the current state of the art is that different works are rarely comparable: they often discuss different parameter counts, training data, and evaluation methodology.This paper proposes a comparison of multilingual pretraining objectives in a controlled methodological environment. We ensure that training data and model architectures are comparable, and discuss the downstream performances across 6 languages that we observe in probing and fine-tuning scenarios.We make two key observations: (1) the architecture dictates which pretraining objective is optimal; (2) multilingual translation is a very effective pretraining objective under the right conditions.We make our code, data, and model weights available at https://github.com/Helsinki-NLP/lm-vs-mt.

Les modèles de langue préentraînés (PLM) constituent aujourd’hui de facto l’épine dorsale de la plupart des systèmes de traitement automatique des langues. Dans cet article, nous présentons Jargon, une famille de PLMs pour des domaines spécialisés du français, en nous focalisant sur trois domaines : la parole transcrite, le domaine clinique / biomédical, et le domaine juridique. Nous utilisons une architecture de transformeur basée sur des méthodes computationnellement efficaces(LinFormer) puisque ces domaines impliquent souvent le traitement de longs documents. Nous évaluons et comparons nos modèles à des modèles de l’état de l’art sur un ensemble varié de tâches et de corpus d’évaluation, dont certains sont introduits dans notre article. Nous rassemblons les jeux de données dans un nouveau référentiel d’évaluation en langue française pour ces trois domaines. Nous comparons également diverses configurations d’entraînement : préentraînement prolongé en apprentissage autosupervisé sur les données spécialisées, préentraînement à partir de zéro, ainsi que préentraînement mono et multi-domaines. Nos expérimentations approfondies dans des domaines spécialisés montrent qu’il est possible d’atteindre des performances compétitives en aval, même lors d’un préentraînement avec le mécanisme d’attention approximatif de LinFormer. Pour une reproductibilité totale, nous publions les modèles et les données de préentraînement, ainsi que les corpus utilisés.

La génération de textes neuronaux fait l’objet d’une grande attention avec la publication de nouveaux outils tels que ChatGPT. La principale raison en est que la qualité du texte généré automatiquement peut être attribuée à un$cdot$e rédacteurice humain$cdot$e même quand l’évaluation est faite par un humain. Dans cet article, nous proposons un nouveau corpus en français et en anglais pour la tâche d’identification de textes générés automatiquement et nous menons une étude sur la façon dont les humains perçoivent ce texte. Nos résultats montrent, comme les travaux antérieurs à l’ère de ChatGPT, que les textes générés par des outils tels que ChatGPT partagent certaines caractéristiques communes mais qu’ils ne sont pas clairement identifiables, ce qui génère des perceptions différentes de ces textes par l’humain.

This article presents MedDialog-FR, a large publicly available corpus of French medical conversations for the medical domain. Motivated by the lack of French dialogue corpora for data-driven dialogue systems and the paucity of available information related to women’s intimate health, we introduce an annotated corpus of question-and-answer dialogues between a real patient and a real doctor concerning women’s intimate health. The corpus is composed of about 20,000 dialogues automatically translated from the English version of MedDialog-EN. The corpus test set is composed of 1,400 dialogues that have been manually post-edited and annotated with 22 categories from the UMLS ontology. We also fine-tuned state-of-the-art reference models to automatically perform multi-label classification and response generation to give an initial performance benchmark and highlight the difficulty of the tasks.

Multilingual pretraining and fine-tuning have remarkably succeeded in various natural language processing tasks. Transferring representations from one language to another is especially crucial for cross-lingual learning. One can expect machine translation objectives to be well suited to fostering such capabilities, as they involve the explicit alignment of semantically equivalent sentences from different languages. This paper investigates the potential benefits of employing machine translation as a continued training objective to enhance language representation learning, bridging multilingual pretraining and cross-lingual applications. We study this question through two lenses: a quantitative evaluation of the performance of existing models and an analysis of their latent representations. Our results show that, contrary to expectations, machine translation as the continued training fails to enhance cross-lingual representation learning in multiple cross-lingual natural language understanding tasks. We conclude that explicit sentence-level alignment in the cross-lingual scenario is detrimental to cross-lingual transfer pretraining, which has important implications for future cross-lingual transfer studies. We furthermore provide evidence through similarity measures and investigation of parameters that this lack of positive influence is due to output separability—which we argue is of use for machine translation but detrimental elsewhere.

Pretrained Language Models (PLMs) are the de facto backbone of most state-of-the-art NLP systems. In this paper, we introduce a family of domain-specific pretrained PLMs for French, focusing on three important domains: transcribed speech, medicine, and law. We use a transformer architecture based on efficient methods (LinFormer) to maximise their utility, since these domains often involve processing long documents. We evaluate and compare our models to state-of-the-art models on a diverse set of tasks and datasets, some of which are introduced in this paper. We gather the datasets into a new French-language evaluation benchmark for these three domains. We also compare various training configurations: continued pretraining, pretraining from scratch, as well as single- and multi-domain pretraining. Extensive domain-specific experiments show that it is possible to attain competitive downstream performance even when pre-training with the approximative LinFormer attention mechanism. For full reproducibility, we release the models and pretraining data, as well as contributed datasets.

Neural text generation is receiving broad attention with the publication of new tools such as ChatGPT. The main reason for that is that the achieved quality of the generated text may be attributed to a human writer by the naked eye of a human evaluator. In this paper, we propose a new corpus in French and English for the task of recognising automatically generated texts and we conduct a study of how humans perceive the text. Our results show, as previous work before the ChatGPT era, that the generated texts by tools such as ChatGPT share some common characteristics but they are not clearly identifiable which generates different perceptions of these texts.

This paper presents the results of the SHROOM, a shared task focused on detecting hallucinations: outputs from natural language generation (NLG) systems that are fluent, yet inaccurate. Such cases of overgeneration put in jeopardy many NLG applications, where correctness is often mission-critical. The shared task was conducted with a newly constructed dataset of 4000 model outputs labeled by 5 annotators each, spanning 3 NLP tasks: machine translation, paraphrase generation and definition modeling.The shared task was tackled by a total of 58 different users grouped in 42 teams, out of which 26 elected to write a system description paper; collectively, they submitted over 300 prediction sets on both tracks of the shared task. We observe a number of key trends in how this approach was tackled—many participants rely on a handful of model, and often rely either on synthetic data for fine-tuning or zero-shot prompting strategies. While a majority of the teams did outperform our proposed baseline system, the performances of top-scoring systems are still consistent with a random handling of the more challenging items.

Definition Modeling, the task of generating definitions, was first proposed as a means to evaluate the semantic quality of word embeddings—a coherent lexical semantic representations of a word in context should contain all the information necessary to generate its definition. The relative novelty of this task entails that we do not know which factors are actually relied upon by a Definition Modeling system. In this paper, we present evidence that the task may not involve as much semantics as one might expect: we show how an earlier model from the literature is both rather insensitive to semantic aspects such as explicit polysemy, as well as reliant on formal similarities between headwords and words occurring in its glosses, casting doubt on the validity of the task as a means to evaluate embeddings.

Language models have become a key step to achieve state-of-the art results in many different Natural Language Processing (NLP) tasks. Leveraging the huge amount of unlabeled texts nowadays available, they provide an efficient way to pre-train continuous word representations that can be fine-tuned for a downstream task, along with their contextualization at the sentence level. This has been widely demonstrated for English using contextualized representations (Dai and Le, 2015; Peters et al., 2018; Howard and Ruder, 2018; Radford et al., 2018; Devlin et al., 2019; Yang et al., 2019b). In this paper, we introduce and share FlauBERT, a model learned on a very large and heterogeneous French corpus. Models of different sizes are trained using the new CNRS (French National Centre for Scientific Research) Jean Zay supercomputer. We apply our French language models to diverse NLP tasks (text classification, paraphrasing, natural language inference, parsing, word sense disambiguation) and show that most of the time they outperform other pre-training approaches. Different versions of FlauBERT as well as a unified evaluation protocol for the downstream tasks, called FLUE (French Language Understanding Evaluation), are shared to the research community for further reproducible experiments in French NLP.

French, as many languages, lacks semantically annotated corpus data. Our aim is to provide the linguistic and NLP research communities with a gold standard sense-annotated corpus of French, using WordNet Unique Beginners as semantic tags, thus allowing for interoperability. In this paper, we report on the first phase of the project, which focused on the annotation of common nouns. The resulting dataset consists of more than 12,000 French noun occurrences which were annotated in double blind and adjudicated according to a carefully redefined set of supersenses. The resource is released online under a Creative Commons Licence.

Les modèles de langue pré-entraînés sont désormais indispensables pour obtenir des résultats à l’état-de-l’art dans de nombreuses tâches du TALN. Tirant avantage de l’énorme quantité de textes bruts disponibles, ils permettent d’extraire des représentations continues des mots, contextualisées au niveau de la phrase. L’efficacité de ces représentations pour résoudre plusieurs tâches de TALN a été démontrée récemment pour l’anglais. Dans cet article, nous présentons et partageons FlauBERT, un ensemble de modèles appris sur un corpus français hétérogène et de taille importante. Des modèles de complexité différente sont entraînés à l’aide du nouveau supercalculateur Jean Zay du CNRS. Nous évaluons nos modèles de langue sur diverses tâches en français (classification de textes, paraphrase, inférence en langage naturel, analyse syntaxique, désambiguïsation automatique) et montrons qu’ils surpassent souvent les autres approches sur le référentiel d’évaluation FLUE également présenté ici.

Les schémas Winograd sont des problèmes de résolution d’anaphores conçus pour nécessiter un raisonnement sur des connaissances du monde. Par construction, ils sont insensibles à des statistiques simples (co-occurrences en corpus). Pourtant, aujourd’hui, les systèmes état de l’art pour l’anglais se basent sur des modèles de langue pour résoudre les schémas (Trinh & Le, 2018). Nous présentons dans cet article une étude visant à tester des modèles similaires sur les schémas en français. Cela nous conduit à revenir sur les métriques d’évaluation utilisées dans la communauté pour les schémas Winograd. Les performances que nous obtenons, surtout comparées à celles de Amsili & Seminck (2017b), suggèrent que l’approche par modèle de langue des schémas Winograd reste limitée, sans doute en partie à cause du fait que les modèles de langue encodent très difficilement le genre de raisonnement nécessaire à la résolution des schémas Winograd.

As opposed to word sense induction, word sense disambiguation (WSD) has the advantage of us-ing interpretable senses, but requires annotated data, which are quite rare for most languages except English (Miller et al. 1993; Fellbaum, 1998). In this paper, we investigate which strategy to adopt to achieve WSD for languages lacking data that was annotated specifically for the task, focusing on the particular case of verb disambiguation in French. We first study the usability of Eurosense (Bovi et al. 2017) , a multilingual corpus extracted from Europarl (Kohen, 2005) and automatically annotated with BabelNet (Navigli and Ponzetto, 2010) senses. Such a resource opened up the way to supervised and semi-supervised WSD for resourceless languages like French. While this perspective looked promising, our evaluation on French verbs was inconclusive and showed the annotated senses’ quality was not sufficient for supervised WSD on French verbs. Instead, we propose to use Wiktionary, a collaboratively edited, multilingual online dictionary, as a resource for WSD. Wiktionary provides both sense inventory and manually sense tagged examples which can be used to train supervised and semi-supervised WSD systems. Yet, because senses’ distribution differ in lexicographic examples found in Wiktionary with respect to natural text, we then focus on studying the impact on WSD of the training data size and senses’ distribution. Using state-of-the art semi-supervised systems, we report experiments of Wiktionary-based WSD for French verbs, evaluated on FrenchSemEval (FSE), a new dataset of French verbs manually annotated with wiktionary senses.