ANR InExtenso:

WP4 of the project is dedicated to the evaluation of the environmental impact of the LLMs. More precisely, it aims at proposing a method for measuring the environmental impact of digital health and use it in the project evaluations and beyond.

7.1.1 ACGtk

• Name:
  Abstract Categorial Grammar Development Toolkit
• Keywords:
  Natural language processing, Functional programming, Logic programming, Lambda-calculus, Ocaml
• Scientific Description:

  Abstract Categorial Grammars (ACG) are a grammatical formalism in which grammars are based on typed lambda-calculus. A grammar generates two languages: the abstract language (the language of parse structures), and the object language (the language of the surface forms, e.g., strings, or higher-order logical formulas), which is the realization of the abstract language.

  ACGtk provides two software tools to develop and to use ACGs: acgc, which is a grammar compiler, and acg, which is an interpreter of a command language that allows one, in particular, to parse and realize terms.

• Functional Description:
  ACGtk provides a piece of software for developing and using Abstract Categorial Grammars (ACG).
• Release Contributions:
  This new version of the software provides heavy internal modifications to manage and enumerate the parses of an expression, in particular when the latter is highly ambiguous. These modifications also prepare the integration of features to sort the results according to weight information associated with grammatical rules.
• URL:
  https://­gitlab.­inria.­fr/­ACG/­dev/­ACGtk
• Publications:
  hal-01242154, hal-01328702, tel-01412765, inria-00112956, inria-00100529, hal-04479621
• Contact:
  Sylvain Pogodalla
• Participants:
  Philippe De Groote, Pierre Ludmann, Jiri Marsik, Sylvain Pogodalla, Vincent Tourneur

7.1.2 Grew

• Name:
  Graph Rewriting
• Keywords:
  Semantics, Syntactic analysis, NLP, Graph rewriting
• Functional Description:
  Grew is a Graph Rewriting tool dedicated to applications in NLP. Grew takes into account confluent and non-confluent graph rewriting and it includes several mechanisms that help to use graph rewriting in the context of NLP applications (built-in notion of feature structures, parametrization of rules with lexical information).
• News of the Year:
  In 2024, one new version 1.16 was released (together with several bug fixes). The major new features are: new code for corpusbank handling and new syntax in request and key clustering (`length` and `delta` between two nodes)
• URL:
  https://­grew.­fr/
• Publications:
  hal-03724068, hal-03177701, hal-01930591, hal-01814386, hal-03021720, hal-04387830, hal-04387852, hal-03724068, hal-03724129, hal-03846825
• Contact:
  Bruno Guillaume
• Participants:
  Bruno Guillaume, Guy Perrier, Guillaume Bonfante

7.1.3 HostoMytho

• Keywords:
  Game with a purpose, Natural language processing
• Functional Description:
  HostoMytho is a GWAP, or "game with a purpose" developed within the framework of the CODEINE ANR project. The aim of the game is to allow users to annotate medical files generated automatically, in order to evaluate their plausibility (quality of the language and medical semantics) and to add different layers of information (negation, hypothesis, time, etc.). HostoMytho is multiplatform.
• URL:
  https://­hostomytho.­atilf.­fr
• Publication:
  hal-04555052v1
• Contact:
  Karën Fort
• Partners:
  LISN, CEA-List

7.1.4 Arborator-Grew

• Name:
  Arborator's Collaborative Annotation
• Keywords:
  Annotation tool, Syntactic analysis
• Functional Description:
  The online interface allows managing collaborative annotation projects in dependency syntax. It is possible to use Grew queries and also to directly rewrite graphs in the annotation tool.
• News of the Year:
  During 2024, we redesigned the tool's interface, refactored the code base for both frontend and backend. In parallel, we set up pipelines for automatic deployment and improved code documentation. In addition, we continued to work on improving existing functionalities and adding new ones based on user requests.
• URL:
  https://­arborator.­grew.­fr/
• Publication:
  hal-03021720
• Contact:
  Bruno Guillaume
• Participants:
  Khensa Amani Daoudi, Bruno Guillaume, Gael Guibon, Kim Gerdes, Kirian Guiller
• Partners:
  Université Paris Nanterre, LIMSI, LISN

8.1.1 Abstract Categorial Grammars

8.1.2 Formal semantics of dependency relations

We have continued to work on our formal semantic theory of dependency grammars. This theory, which is fully compositional and robust, is based on the following principles:

1. Dependency relations are represented as binary functions of type $\alpha \to \beta \to \alpha$ (or $\beta \to \alpha \to \alpha$), where $\alpha$ is the syntactic category of the governor and $\beta$ is the syntactic category of the governee.
2. By virtue of a coherence principle, the different ways of encoding a dependency structure by means of a $\lambda$-term should all give rise to the same semantic interpretation.
3. The semantic interpretation of a syntactic category cat is a type of the form ${\beta }_1\to \cdots {\beta }_n\to \alpha$, where $\alpha$ is the Montagovian interpretation of cat and ${\beta }_1,...{\beta }_n$ are the Montagovian interpretations of the syntactic categories of the phrases whose heads can potentially be governed by the head of a phrase of category cat.
4. Saturating operators allow phrases to recover their usual Montagovian interpretations.
5. Verbs and verbal phrases are semantically interpreted as sets of sets of events.

We have tested these principles on advanced syntactic phenomena 45. We considered two cases: the relative clauses (which depend on the acl:relcl dependency relation) and the open clausal complements (which depend on the xcomp dependency relation).

8.1.3 Formal semantics of adnominal modification

We have proposed a treatment of adnominal modification that parallels the treatment of adverbial modification in neo-Davidsonian event semantics 42. To this end, we introduced a notion of perspective that allows nouns to be interpreted as sets of sets of perspective. The resulting theory provides a unified compositional treatment of intersective, subsective, modal, and privative adjectives, and avoids the intensional paradoxes caused by an extensional treatment of subsecutive adjectives.

8.1.4 Semantic Representation

In 27, we presented the first version of YARN, a new semantic representation formalism which aim is to unify the advantages of logic-based formalisms while retaining direct interpretation, making it widely usable. YARN is rooted in the encoding of different semantic phenomena as separate layers. The paper presents a formal definition of the mathematical structure that constitutes YARN and illustrates with concrete examples how this structure can be used in the context of semantic representation for encoding multiple phenomena (such as modality, negation and quantification) as layers built on top of a central predicate-argument structure. The benefit of YARN is that it allows for the independent annotation and analysis of different phenomena as they are easy to “switch off”. Furthermore, the ability of YARN to encode simple interactions between phenomena is explored. The authors wrap up the work presented by a discussion of some of the interesting observations made during the development of YARN so far and outline our extensive future plans for this formalism.

Zhengjian Li conducted an M1 internship under the supervision of Siyana Pavlova, Bruno Guillaume and Maxime Amblard. He studied the semantic annotation of real data in Yarn and developed an interface to automatically produce Yarn representation graphs.

As part of a mobility program at Sapienza University under the supervision of Roberto Navigli, Hee-Soo Choi collaborated with the SapienzaNLP team on a multilingual semantic parsing project proposing a strategy to semi-automatically generate semantic annotations across languages from LLMs. The work was published at the international conference ACL 2024 25.

8.1.5 Syntax and semantics of questions

Natural language statements are composed not only of declarative sentences but also of interrogative ones. Moreover, sentences cannot be categorized into purely declarative or purely interrogative sentences. Typically, a declarative statement may contain a subordinated interrogative clause:

• (a)
  I don't know where Mary is.

In 28, a kind of French subordinated interrogative which has not been observed before is described and analyzed. It consists of an adverbial modifier clause introduced by a preposition, e.g. selon (`depending on'):

• (b)
  Selon comment vous vous positionnez, vous n'aurez pas tous la même perception.

Up to now, subordinated interrogative clauses have been claimed to only be acceptable as subjects or complements of verbs, nouns or adjectives. This discovery widens our comprehension of the potential contribution of interrogatives in discourse.

This interaction between declarative and interrogative clauses is particularly present in dialogues, where the logical notion of answerhood is as significant as the one of inference. In order to tackle this issue from a formal standpoint, we investigated the properties and possible uses of inquisitive semantics. Inquisitive semantics is a formal semantic theory based on a logic that provides a uniform treatment of both declarative and interrogative expressions.

In 29, the interaction of questions and possibility modal operators is investigated. It is shown that when an existential modal scopes over the question operator, several dynamic properties of questions are weakened or blocked. In particular, modalized singular-which-questions have a weaker exhaustivity, uniqueness presuppositions and block anaphoric reference to the wh-word, except in modal subordination. A model based on dynamic inquisitive semantics and capturing these phenomena is provided.

8.1.6 Use of semantics

Before the invention of the printing press, texts could only be reproduced through manual copying, a process prone to errors, accidents, and intentional modifications. These changes altered each manuscript and were subsequently propagated by other scribes. For philologists reconstructing text history and genealogical relationships (stemma codicum), analyzing these variants is crucial. Stemmatology methods aim to objectively construct genealogical trees of textual transmission.

At the University of Lorraine, the Écritures laboratory and MSH have focused on uncovering the genealogical lineage of Hebrew manuscripts. A join project with Maxime Amblard seeks to improve the manual work involved in critical editions of the Hebrew Bible by applying advanced methods from applied mathematics and natural language processing to reconstruct stemmas.

Iglika Zlatkova Nikolova-Stoupak is recruted to develop stemmatology algorithms. She design, train and test learning model to automatically tag scribal variants in manuscripts. The model will classify variants (orthographic, lexical, grammatical, etc.) based on expert-provided data and suggest classifications for new string comparisons. She also develop methods to compute semantic-based distances between Hebrew words, enabling the comparison of variant meanings. This will involve creating textual embeddings and neural network-based representations of Hebrew words.

Maxime Amblard started a collaboration with the French Company Namkin. The industry faces numerous challenges that necessitate the evolution of BtoB marketing tools, in order to develop a valuable offer and provide an enhanced customer experience. Namkin's BrainLab develops industrial marketing tools for digitalizing customer relations, evolving business models, and exploiting business and economic data for business development. One of the key challenges of marketing intelligence is to identify risks and opportunities so as to guide marketing strategies. Among the sources of information useful to detect risks and opportunities, Namkin has identified Business Events, that is, “textually reported real-world occurrences, actions, relations, and situations involving companies and firms”  62.

While modern semantic representations may contain vast quantities of information, they do not always (or necessarily) contain the information that is useful for the concrete application. For instance, significant challenges still persist in dealing with temporal relations and finely-grained negation interpretation.

Recent research has looked into the benefits of exploiting semantic representations, and in particular Abstract Meaning Representation, for low-resources scenarios and document level event argument extraction. However, it appears that AMR has to be adapted in order to optimally support event extraction related tasks  78. One major limitation of AMR for document-level event extraction is that AMR works at the sentence level, and thus requires the aggregation of sentence-level representations. AMR is also limited in terms of negation and universal quantification expressive power.

8.2.1 Dialogue Modeling

Discourse relation prediction arguably is the most difficult task in discourse parsing. Previous studies have generally focused on explicit or implicit discourse relation classification in monologues, leaving dialogue an under-explored domain. Facing the data scarcity issue, Chloé Braud (IRIT), Maxime Amblard and Chuyuan Li proposed to leverage self-training strategies based on a Transformer backbone. Moreover, they designed the first semi-supervised pipeline that sequentially predicts discourse structures and relations. Using 50 examples, their relation prediction module achieves 58.4 in accuracy on the STAC corpus, close to supervised state-of-the-art. Full parsing results show notable improvements compared to the supervised models both in-domain (gaming) and cross-domain (technical chat), with better stability 24.

Maxime Amblard and Amandine Decker continue to work on topic segmentation 17 and topical structure modeling 18. Topics play an important role in dialogue coherence, as what is currently discussed constrains the possible contributions of the participants, and initiating a topic while the previous one is still under discussion may be confusing without appropriate signals. However, the way to actually define the notion of topic is debated in linguistics and not sufficiently discussed in dialogue modeling. A precise description of topics and topic shifts in conversation would contribute to a better understanding of what makes us judge a sequence of utterances to be coherent.

In 17, linear topic segmentation is performed on transcriptions of the Friends TV show. These contain multi-party conversations in a daily life setting which are complex to study due to their chaotic nature. Linear topic segmentation proves to be ill-suited for such complicated dialogues. Hence a new topic segmentation method must be investigated.

36 presents a method to build an unsupervised topic similarity measure. It relies on the existing tree structure of Reddit threads and allows us to analyse the evolution of topics over the course of a conversation.

Ines Hernandez conducted an M2 internship under the supervision of Amandine Decker and Maxime Amblard. Her study began by enhancing a topic similarity model. Since topicality is a key feature while differentiating subordinating and coordinating relations, she then applied this improved model for the specific task of classifying these relations using a conversational dataset. To test its performance she utilized three different datasets which contain information from different sources and have different structures, in English and French. Additionally, she proposed an analysis of the results by rhetorical relations, in the attempt to observe if certain rhetorical relations are more straight-forward to classify. Furthermore, she fine-tunes pre-train language models, such as BERT, roBERTa, deBERTa, and Llama2 to continue the experiments on this classification task. Notably, BERT outperforms other models in classifying these relations, possibly due to its Next Sentence Prediction (NSP) training objective. Finally, she proposes to provide context to the models during training, since this is considered crucial for understanding how segments are connected. This shows potential results, improving in particular the BERT model.

In order to analyze different types of topic shifts, 18 propose to create a corpus of written task-oriented conversations (discussion of the ethical dilemma of the balloon task), where the dialogues happen by message exchanges. Such a controlled setting where the main topic is fixed, and subtopics are more easily identifiable, could be very helpful when it comes to understanding how people change the topic and react to topic shifts in dialogues. Emeric Licorni conducted an L3 internship under the supervision of Amandine Decker and Maxime Amblard to explore the design of this experiment. Applying the topic shift similarity metric mentioned above to human annotated conversations could enable us to develop a hierarchical topic segmentation model.

8.2.2 Discourse Markers

Jacques Jayez is currently working with Mathilde Dargnat (ATILF), Paola Herreño (Ph.D. candidate ATILF-LLF) and Maeva Sillaire (Ph.D. candidate ATILF) on the semantic representation of D(iscourse) M(arkers). DM are words/expressions like so or well in English which help structuring discourse or communicating speakers' internal epistemic or affective states. The domain-based approach initiated in the 90s consists in defining different types (aka domains) of semantic objects, like states of affairs, beliefs or speech acts. Such objects have a rich internal structure, which calls for a sufficiently expressive representation. For instance, speech acts involve discourse planning. Moreover DM can impact various layers of meaning (propositional content, presupposed content, etc.). Our goal is to implement formally the domain intuition by using tools from languages with a flexible subtyping mechanism (Ocaml) and to investigate whether the notion of monad, familiar in Haskell, can help us to characterize some `side effects' of DM.

At the moment (January 2025), we have constructed a libray of about 600 automata in the language of the Unitex-Gramlab software, to detect occurrences of about 700 DM. For words or expressions which are lexically ambiguous, this tools helps us to identify their uses as DM. For instance, bon can be an adjective (un bon roman, a good novel) or a discourse particle like in bon, il faut qu'on parte (OK, we have to go). A comparison with the results of various taggers (probabilistic and based on LLM) is under way. We have also attacked the typing problem mentioned above by (i) clarifying the different distinctions necessary to have a clear notion of type for DM and (ii) by capturing the basic information structures for DM in the language of Dialogue Game Boards of 56 (64).

Jacques Jayez is also working on the argumentative dimension of discourse, using a combination of standard Bayesian approach and game-theoretical notions 23, 63.

8.2.3 Pathological Discourse Modeling

Michel Musiol is once again part-time delegate at Inria (délégation nationale SHS) for the period 2023-2024. This proximity has enabled us to pursue an active collaboration on the modeling of pathological and clinical discourse.

In the context of the MePheSTO project (Digital Phenotyping for Psychiatric Disorders from Social Interaction - DFKI-Inria AI project), a multimodal perspective on this issue has been introduced. This includes the development of an interlocutory model of what might be termed a "therapeutic effect" in clinical/psychopathological interviewing, with regard to the cognitive-discursive profile and oculomotor behavior of participants (Amandine Lecomte ś thesis forthcoming 2025; Marie-Hélène Pierre's thesis forthcoming 2025).

Indeed, based on a previous study addressing the issue of cognitive impairment bypass based on the dynamics of the repetition process 65, 66 and another modeling visual attention in the psychologist-schizophrenic patient interaction 73 as well as on the basis of new data (for example, from subjects with ultra-high risk of developing psychosis or subjects who have experienced a major depressive episode 55, in the MePheSTO project), Maxime Amblard, Michel Musiol and Amandine Lecomte have explored the dependent relationships between supportive visual behavior and interaction dynamics, on the one hand, and the relationships between mental pathology and oculomotor disorders, on the other hand.

In addition, in 70 they have sketched a methodology for analyzing speech disorders, which will have the particularity of helping to select the discontinuous sequences most likely to carry thought disorders. They have anticipated the development of a modeling system based on the principles of pragmatic linguistics and formal semantics, which, when applied to carefully selected discontinuous discourse sequences 69, will have a good chance of revealing the nature of the underlying thought disorders. They compared the conjectures with the results of an earlier study on the discovery of four "proven" types of discontinuous sequences, and showed which of these sequences can therefore be considered carrying thought disorders. All the area of thought disorders in schizophrenic disourse might be modelized on the basis of computational and semantical tools 72, 75.

They also analyzed these sequences by testing certain principles of semantic modeling in order to identify the nature of the disorders and thought operations underlying the discontinuous sequences concerned 71, 70. They show that discursive thought disorders should not be considered simply as the expression of a dysexecutive syndrome 52, but also as a device likely to affect more complex thought operations such as inferences involved in the conversational context representation system, in semantic memory and in calculating the meaning of utterances or in calculating the meaning of the speaker.

Improving the heuristics of formal systems for recognizing speech disorders and interpreting thought disorders on the basis of more appropriate and accurate semantic modeling may lead to the development of more discriminating and effective diagnostic tools, particulary to account for Schwachwan-diamond diagnostic and cognitive dysfunction 77 The formatting of the formal systems they have achieved will make it possible to represent the interlocutory structure of the disorder more and more accurately in its natural context of expression (speech), and should lead to the development of computerized diagnostic aids.

Finally, the increased precision of formal modeling applied to communication disorders should also make it possible to test the hypothesis that certain discourse configurations are related to thought disorders (in the broad sense), while others reveal cognitive dysfunctions that have more to do with the conditions of the possibility of discourse.

In this line, Michel Musiol has also updated MOS-SF36 norms in the young French population 76. This program focuses on accurate and rapid diagnosis, as well as long-term therapeutic follow-up. These are major challenges for contemporary psychopathology. Thanks to modeling 65, 66 and computing 51, this work is developing multimodal methodologies for investigating symptoms (language, speech, neuropsychological and cognitive processes, eye movements and visual attention) that are sufficiently accurate to give rise in the medium term to the development of computerized diagnostic and therapeutic follow-up tools for the benefit of those involved in mental health.

Improving the accuracy of these tools also requires to refine the aimed targets (i.e. the definition of the symptoms). In this line, Vincent Martin has been involved in a work aiming at unifying the semiology of psychiatry through data vizualisation 13.

Rémi De Vergnette conducted an M1 intersnship under the supervision of Maxime Amblard. The internship focused on developing methods for the automatic analysis of interview data. Using video recordings and eye-tracker data, the goal was to extract and utilize a list of fixations made by the psychologist and patient, along with the corresponding fixation zones. Previously, much of this work was performed manually using commercial analysis software. Automating this process was essential due to the large number of interviews to analyze (150 across two hospitals in Nice and Aix-en-Provence), with further data collections planned. Reducing manual processing allows more time to focus on in-depth analyses requiring expert input from psychologists. He developed several important works listed afterward. The development of an algorithm for zone-based video segmentation. The creation of algorithms to associate fixation data with zones, accounting for biases introduced by the software and uncertainties in measurement values. The implementation of a module to extract raw eye-tracker data from acquisition files. The design of two interfaces to facilitate collaboration with participating psychologists.

8.2.4 Cognitive traces of side issues

It is by now widely believed that natural language communication operates at several levels. This means that information is distributed across several partially independent dimensions. For instance, a sentence like My stupid neighbor made noise again, simultaneously conveys that my neighbor made noise (the truth-conditional content), that the speaker considers he is stupid (an expressive, side issue 1) and that he had made noise before (the presupposition of again, side issue 2). While these phenomena have been extensively described from an empirical perspective, there is at present no unified framework for representing their differences and possible interactions under a formal, computational or cognitive point of view.

We examined the motor effects of presuppositions, using the convenient lexical material of factive verbs, that is, verbs which presuppose the truth of the complement clause. For example, Mary knows that Paul cheated on the exam presupposes that Paul cheated on the exam (side issue) and asserts (truth-conditional content) that she believes that. It has been shown that the oral presentation of movement-related verbs like jump or push elicits some activation in the motor cortex and finally results into an involuntary contraction of the thumb-index arc, which can be recorded by a special electromagnetic cell, called a grip force sensor.

We adapted this technique to the case of factive verbs on a series of sentences of the form Mary knows that Paul throws the ball, compared with high base-level sentences like Paul throws the ball and low base-level sentences like Paul does not throw the ball. Summarizing, our results indicate that the sentences with the factive verbs elicit a very similar response to that of high base-level ones, and, as expected, a different response from that of low base-level one. This suggests that, at least for factive verbs, the presupposed status leaves no trace of a special cognitive treatment, which would lead for instance, to a delayed or weaker motor response.

However, when applied to more complex negative sentences like Mary does not know that Paul throws the ball, there is no evidence of a motor trace. This is in agreement with observations in the literature suggesting that, under negation, presuppositions are more 'difficult' to process than in simple assertive sentences. More precisely, in the case of motor response, negation interacts with the presupposition, which suggests that truth-conditional content and side issue cognitive processing cannot be totally separated 9.

8.3.1 Universal Dependencies and Surface Syntactic Universal Dependencies

The Universal Dependencies project (UD) aims at building a syntactic dependency scheme which allows for similar analyses for several different languages. Bruno Guillaume is active in the UD community, and participate in the development and the improvement of the French data in this international initiative.

During 2024, he continued working, in collaboration with Sylvain Kahane, Kim Gerdes and their teams on the promotion of the Surface Syntactic Universal Dependencies (SUD) framework. SUD is an annotation scheme for syntactic dependency treebanks, that is almost isomorphic to UD (Universal Dependencies). Contrary to UD, it is based on syntactic criteria (favoring functional heads) and the relations are defined on distributional and functional bases.

This work is mainly conducted in the ANR project Autogramm (Induction of descriptive grammar from annotated corpora) started in 01 2022. The goal of this project is to automate, as far as possible, the extraction of descriptive grammars and grammatical descriptions from annotated corpora for linguistic and typological studies. The project also promotes the development of treebanks for low-resourced languages, in order to extract quantitative descriptive grammars for these languages.

Bruno Guillaume, in collaboration with Althea Löfgren, Santiago Herrera, Sylvain Kahane and Natalia Levshina have conducted a comparative analysis of UD Treebanks to explore how they can asset for linguistic diversity in linguistic typology. This work was presented in 43.

In 2024, two new versions of Universal Dependencies were released. Bruno Guillaume worked in collaboration with field linguists to produce or enhanced treebanks in Surface Syntactic Universal Dependencies and to convert these treebanks to Universal Dependencies:

• Version 2.14 on May:
  • new treebank for Northern Hausa (with Bernard Caron)
  • new treebank for Southern Hausa (with Bernard Caron)
• Version 2.15 on November:
  • enhanced treebank for Beja (with Martine Vanhove)
  • new treebank for Gbaya (with Paulette Roulon)
  • new treebank for Pesh (with Natalia Cáceres Arandia)

In 20, different ways to annotate both syntactic and morphological relations in a dependency treebank are compared. New formats, called mSUD and mUD, are proposed; they are compatible with the Universal Dependencies (UD) schema for syntactic treebanks. The paper emphasizes on mSUD rather than mUD, the former being based on distributional criteria for the choice of the head of any combination, which allows to clearly encode the internal structure of a word, that is, the derivational path. The authors investigate different problems posed by a morph-based annotation, concerning tokenization, choice of the head of a morph combination, relations between morphs, additional features needed, such as the token type differentiating roots and derivational and inflectional affixes. The annotation schema is then applied to different languages from polysynthetic languages such as Yupik to isolating languages such as Chinese.

The two papers 32, 39, present a new phonetic resource for Nigerian Pidgin, a low-resource language of West Africa. Aiming to provide a new tool for research on intonosyntax, the authors have augmented an existing syntactic treebank of Nigerian Pidgin, associating each orthographically transcribed token with a series of syllable-level alignments and phonetizations. Syllables are further described using a set of continuous and discrete prosodic features. This new approach provides a simple tool for researchers to explore the prosodic characteristics of various syntactic phenomena. The papers present the format of the corpus, the various features added, and several explorations that can be performed using an online interface. A prosodically specified lexicon extracted using this resource is also presented. In it, each orthographic form is accompanied by the frequency of its phoneme-level variants, as well as the suprasegmental features that most frequently accompany each syllable. Finally, several additional case studies on how this corpus can used in the study of the language’s prosody are presented.

Bruno Guillaume is a member of the core group of the UniDive COST action where he is in charge of the working groups dedicated to “corpus annotation”. This action was presented in a workshop in 31.

8.3.2 Enriching Lexical Resources

As part of her thesis on French lexical resources, Hee-Soo Choi conducted experiments on the enrichment of lexical semantic graphs using link prediction models. She proposed a resource-centric approach of the link prediction task by generating confidence-aware predictions that can complete a sparse lexical semantic graph. This work was published at the international conference LREC-COLING 2024 16 and its French version at the national conference TALN 35.

8.3.3 FENEC

Karën Fort worked with Alice Millour (Université Paris 8) and Yoann Dupont (Université Paris 3) on the creation of a balanced sample corpus for French named entity recognition. The created corpus, FENEC, is freely available on GitHub, and was presented in a paper at TALN 2022. An extended version in English was presented in 2024 at LREC-COLING 26.

8.3.4 Synthetic clinical texts generation

In the context of the CODEINE ANR project and more specifically of Nicolas Hiebel's PhD thesis, Karën Fort worked with Aurélie Névéol (LISN-CNRS) and Olivier Ferret (CEA) on the generation of synthetic clinical texts.

The key idea of the project is to use confidential corpora to automatically generate anonymous synthetic texts capable of emulating real documents from the perspective of their linguistic characteristics. Nicolas Hiebel worked with Hugo Boulanger (another Olivier Ferret PhD student) on generating clinical texts using constraints. This work has been published at the Clinical NLP workshop 15 and in French at TALN 33 and in 38.

Another part of the project consists in using a Games With A Purpose to validate and then annotate the synthesized clinical texts. This game, developed by Bertrand Remy, is called HostoMytho (see Section 7.1.3), and includes various mini-games for different annotation layers, such as negation, error typing, or plausibility rating. The game is multi-platform, and therefore intended to be used on the web (see: online HostoMytho), on Android and iOS. It has been presented at the Games and NLP workshop in 2024 22.

8.4.1 Ethics@Loria

Karën Fort originated a working group at LORIA for AI ethics (ethics@loria), involving researchers from various teams, including Maxime Amblard, Marc Anderson, Armelle Brun (BIRD), Mathieu d'Aquin (Orpailleur), Christophe Cerisara (Synalp), Anne Bonneau (Multispeech), Slim Ouni (Multispeech) and Abdessamad Imine (Pesto). Aurore Coince helped manage the group. Ethics@loria proposed the Doctoral training on Ethics "Write your dystopia" in 2022 .

One major outcome of the training, which took some time to be produced, is a book collecting the six short stories written by the PhD students in three languages (French, English and Spanish) 50. It was translated in English and French and made available both as a book and online in 2024: thinkbeforeloading.loria.fr thanks to many people's (esp. Marie Baron's) efforts.

8.4.2 Evaluating Stereotypes in Masked Language Models in Many Languages

Following the creation of French CrowsPairs 7, Karën Fort contacted researchers interested in creating a CrowsPairs corpus for their language, in order to test the language models. The group got bigger and bigger to finally include 22 researchers (including Fanny Ducel, from the team) for 7 languages (German, Maltese, Spanish, Italian, Chinese, standard Arabic and Catalan). The group worked together for more than a year to produce corpora and test masked language models in these 7 new languages. The corpora are freely available, along with the code to test the language models and the guidelines we followed for the adaptation.1 It's important to notice that this work has been performed without any funding. The work has been detailed in a paper presented at LREC-COLING 2024 19.

8.4.3 Evaluating stereotypes in autoregressive language models

Fanny Ducel authored a critical literature review on the topic of stereotypical biases in language models, under the supervision of Karën Fort and Aurélie Névéol. This work, presented in an earlier version at the Workshop on Algorithmic Injustice in Amsterdam in June 2023, has been published in a longer version in the French journal "Revue TAL" 12.

Fanny Ducel, under the supervision of Karën Fort and Aurélie Névéol, measured gender stereotypical biases in cover letters generated by autoregressive language models, in French and Italian. This work has been presented and published at TALN in French 37, and in the international Language Resources and Evaluation journal 11.

In the context of the "Ethics and NLP" ("Éthique et TAL") day, organized at LORIA by Karën Fort, Aaron Boussidan, Fanny Ducel, Karën Fort and Aurélie Névéol presented an abstract on the use of ChatGPT in research on bias 34. Fanny Ducel also presented some actionable directions for bias research at the Workshop "New Perspectives on Bias and Discrimination in Language Technology" in Amsterdam in November 2024 41.

Karën Fort is PI of a new 4 year ANR project (2023-2027), InExtenso (Intrinsic and Extrinsic evaluation of biases in large language models), in collaboration with Rouen's hospital (CHU) and LISN-CNRS. The project aims at better identifying stereotyped biases in LLMs in French and, when possible, mitigate them. Within the framework of this project, Clémentine Bleuze, under the supervision of Karën Fort and Aurélie Névéol, started a thesis on the perception and evaluation of biases in medical applications of Large Language Models.

8.4.4 NLP for NLP and Ethics

Clémentine Bleuze conducted an M2 internship under the supervision of Karën Fort and Maxime Amblard, and in collaboration with Fanny Ducel. During this internship, she worked on the notion of scientific overclaiming (when researchers inadequately interpret or present elements of their research) in NLP papers. Contributions of this work include the definition of a taxonomy of relevant research claims, the constitution of a corpus of NLP claims originating from ArXiv and ACL papers (a subpart of which is human-annotated), and the training of BERT-based models to predict claim types. This work had led to the presentation of a poster during the scientific days of thematic research network LIFT 2 in October 2024 40.

8.4.5 Ethics in AI Integration into Industry

The ongoing collaboration between Karën Fort and Marc Anderson resulted in a final common article in 2024 10 about how the recommendations made by the ethics group were taken into account in the AI-Proficient project.

9.1.1 Other european programs/initiatives

• Bruno Guillaume is a member of the core group of the COST action: CA21167 - Universality, diversity and idiosyncrasy in language technology (UniDive). He is the leader of the working group named "Corpus Annotation".

9.2.1 ANR Project: InExtenso

Participants: Karën Fort, Maxime Amblard, Michel Musiol, Fanny Ducel.

• Title:
  Intrinsic and Extrinsic evaluation of biases in large language models
• Duration:
  10 2023–09 2027
• Coordinator:
  Karën Fort
• Partners:
  CHU Rouen, LISN, LORIA
• Participants:
  Maxime Amblard, Fanny Ducel, Karën Fort (coordinator), Michel Musiol, Miguel Couceiro
• Abstract:
  Large Language Models (LLM) are the Swiss Army knife of today’s Natural Language Processing (NLP). They often outperform the state-of-the-art on benchmarks commonly used in the field for tasks such as part-of-speech tagging, text classification and named-entity recognition, thus paving the way to a myriad of end-user applications. However, it has been shown that LLM exhibit major ethical issues including significant environmental impact, mirroring and amplification of stereotyped biases, which in turn have a disproportionate impact on historically disadvantaged social groups. It is urgent to address the social impact of NLP as the applications we develop, such as chatGPT, are now directly made available to end users. The detection and mitigation of biases have therefore become an active area of research in the past few years, focusing mainly on Masked Language Models (MLM) such as BERT in English and the North American social context. Several sources of bias were identified in the NLP pipeline. However the interconnection between sources and overall impact of each source on downstream applications remains unclear. In this project, we want to observe the entire pipeline, from the intrinsic point of view (within the model itself), to the pre-training task point of view (in the case of autoregressive LLM, text generation), on to some real-world downstream applications. We chose to focus on two types of medical applications: mental illness diagnosis help and information extraction from clinical records for public health purposes such as patient enrollment into clinical trials. The project will provide corpora and methods for a global evaluation of bias in LLM in French as well as studies to further the understanding of biases in clinical NLP pipelines and the environmental impact of the integration of these models in digital health.

9.2.2 ANR Project: CoDeinE

Participants: Karën Fort, Bruno Guillaume, Bertrand Remy.

• Title:
  artificial text COrpus DEsIgNed Ethically automatic synthesis of clinical documents
• Duration:
  03 2021–02 2026
• Coordinator:
  Aurélie Névéol (Limsi)
• Partners:
  CRC, CEA List, LISN, LORIA
• Participants:
  Bruno Guillaume, Karën Fort (local coordinator), Bertrand Remy
• Abstract:
  Machine learning methods have become prevalent in language technologies. They rely on annotated corpora to train models and evaluate algorithms. The CoDeinE project proposes to address the lack of shareable corpora in sensitive domains such as health or banking. The key idea of the project is to use confidential corpora to automatically generate synthetic texts that mimic the linguistic properties of real documents while preserving confidentiality. We will use clinical documents in electronic patient records as a case study. Furthermore, the project will rely on Games With A Purpose and crowd sourcing to validate and annotate the synthesized texts.

9.2.3 ANR Project: Autogramm

Participants: Bruno Guillaume, Karën Fort, Khensa Amani Daoudi.

• Title:
  Induction of descriptive grammar from annotated corpora
• Duration:
  01 2022–12 2025
• Coordinator:
  Sylvain Kahane (Université Paris Nanterre)
• Partners:
  MoDyCo, LACITO, LISN, Inria Nancy – Grand Est
• Participants:
  Bruno Guillaume (local coordinator), Karën Fort
• Abstract:
  The goal of this project is to automate, as far as possible, the extraction of descriptive grammars and grammatical descriptions from annotated corpora for linguistic and typological studies. The project also promotes the development of treebanks for under-endowed languages, in order to extract quantitative descriptive grammars for these languages. The project uses the annotation scheme SUD (Surface-syntactic Universal Dependencies), the query tool Grew-match and the annotation tool ArboratorGrew.

9.2.4 ANR Project: CODIM

Participants: Maxime Amblard, Jacques Jayez.

• Title:
  Compositionality and discourse markers
• Duration:
  01 2023–12 2026
• Coordinator:
  Mathilde Dargnat (Université de Lorraine and ATILF)
• Partners:
  ATILF, LLF, LORIA
• Participants:
  Maxime Amblard, Jacques Jayez
• Abstract:
  The CODIM project focuses on the two main linguistic resources for organizing monologues or conversations in human languages : D(iscourse) M(arkers)(therefore/donc, well/ben, bon etc. in English/French) and prosody (in particular, intonation). It will evaluate their status with respect to two major views on communication: compositionality (the possibility of combining meaningful expressions into more complex meaningful expressions) and pattern or construction-based approaches (the idea that language users exploit partly `frozen’ strings of words). We will compare the semantic and prosodic properties of simple and complex French DM (e.g. ah + bon) found in corpora for written and spoken French, using a variety of technical tools for DM identification (category-driven text mining), clustering (statistics and Machine Learning) and research in prosody (duration and intensity measures, contour representation). The project fosters a number of collaborations between linguists and computer scientists.

10.1.1 Scientific events: organisation

• Karën Fort: organizer of the Journées éthique et TAL (GDR LIFT, TAL, ATALA), Nancy, 2 4 2024. 46
• Karën Fort: co-organizer, with Sylvain Loiseau (LACITO) and Berthold Crysman (LLF) of the GDR LIFT summer school on Annotation, AnnoDemo. From 2024-06-03 to 2024-06-07.

10.1.2 Journal

• Maxime Amblard: editor in chief of the Revue TAL.

10.1.3 Invited talks

• ACL Teaching NLP Workshop, August 15th, 2024, Karën Fort, Teaching Ethics in Natural Language Processing in practice
• SophIA, Nov. 29th, 2024, Karën Fort, Why we should change the way we evaluate LLMs
• Journée de l'axe Terrains "l'IA en terrains minés", ATILF, June 25th 2024, Karën Fort and Fanny Ducel, Enjeux éthiques de l'IA par le prisme du TAL et des biais stéréotypés
• Festival IA, Avignon (France), Nov. 13th, 2024, Vincent Martin, Table ronde "IA en santé metnale, avancée ou menace pour notre avenir ?"
• Congrès du Sommeil, Lille (France), Nov. 20th, 2024, Vincent Martin, Atelier "Enjeux du recueil sémiologique et d'un vocabulaire standardisé en médecine du sommeil"
• LIFT 2 - Journées de lancement, Orléans (France), Nov. 14th, 2024, Bruno Guillaume, "Enrichissement des banques de dépendances syntaxiques : Deux expériences avec la morphologie et la prosodie"
• Symposium Santée Mentale, January 30th, 2024, Maxime Amblard, ”Apprentissage multitâche pour la détection de la dépression dans les dialogues”
• LIH, Luxembourg, February 5th, 2024, Maxime Amblard and Michel Musiol”NLP & gaze for mental disorders”.
• CERCLES, Maxime Amblard
• ANR Isovote days, Maxime Amblard was invited at the round table ”Intelligence artificielle et traitement de corpus en sciences humaines et sociales : vers une révolution des pratiques scientifiques ?”, October 11th, 2024
• Annual seminar ANR SELEXINI, June 19th, 2024, Hee-Soo Choi, "French Lexical Semantic Graphs: Enrichment by Link Prediction and Integration in a WSD model" 14

10.1.4 Leadership within the scientific community

10.1.5 Scientific expertise

• Sylvain Pogodalla: evaluation for the Inria Quadrant Programme.
• Maxime Amblard: evaluation HCERES du LIASD (Université Paris 8).

10.1.6 Research administration

• Karën Fort
  • Member of the Conseil de Pôle AM2I of Université de Lorraine.
  • Member of CNU 27 (Computer Science): participation to qualifications (then became Professor and had to leave)
• Maxime Amblard
  • Member of CNU 27 (Computer Science)
  • Head of the master in Natural Language Processing (master 1 and 2).
• Vincent Martin
  • Member of the steering committee of the Human Language Technologies college from the French Association for Artificial Intelligence (AFIA).
• Sylvain Pogodalla:
  • Elected member of the comité de centre Inria Nancy – Grand Est.
  • In charge of the local commission IES (information et édition scientifique) of the Inria Nancy – Grand Est and LORIA.
  • Member of the national commission IES of Inria.

10.2.1 Teaching

• Licence:
  • Maxime Amblard, AI Introduction, 14h, L1, Université de Lorraine, France.
  • Maxime Amblard, Chuyuan Li NLP for beginners, 20h, L2, Université de Lorraine, France.
  • Maxime Amblard, Discover Data Processing, 15h, L2, Université de Lorraine, France.
  • Maxime Amblard and Chuyuan Li, Linguistic engineering, 20h, L3, Université de Lorraine, France.
  • Maxime Amblard, Ethical aspects of NLP, 10h, L3, Université de Lorraine, France.
  • Maxime Amblard, Clémentine Bleuze, Découverte du Traitement des Données Langagières, 30h, L2, Université de Lorraine, France
  • Hee-Soo Choi, Advanced Databases, 20h, L2, Université de Lorraine, France.
  • Hee-Soo Choi, Advanced Object-Oriented Programming in Python, 37.5h, L3 NLP major, Université de Lorraine, France.
  • Marie Cousin, Programmation Orientée Objet Avancée, 20h, L2, Université de Lorraine, France.
  • Amandine Decker, Introduction à la Programmation (Java), 18h, L1, Université de Lorraine, France.
  • Karën Fort, Bases de données relationnelles, 55h, L3, Sorbonne Université, France.
  • Karën Fort, Relational databases, 55h, L3, Sorbonne Université, France.
  • Siyana Pavlova, Traitement Automatique des Langues, 10h, L2, Université de Lorraine, France.
  • Siyana Pavlova, Formalismes de représentation et de raisonnement, 30h, L3, Université de Lorraine, France.
  • Siyana Pavlova, Ingénierie des langues, 20h, L3, Université de Lorraine, France.
  • Vincent Tourneur, Administration UNIX, 24h, L2, IUT Charlemagne, Université de Lorraine, France.
  • Vincent Tourneur, Structures de données, 40h, L1, IUT Charlemagne, Université de Lorraine, France.
• Master:
  • Maxime Amblard and Amandine Decker, Methods for NLP, 36h, M1 NLP (IDMC), Université de Lorraine, France.
  • Maxime Amblard, NLP project, 20h, M1 NLP (IDMC), Université de Lorraine, France.
  • Maxime Amblard and Amandine Decker, Dialogue ChatBot and Question Answering, 21h, M2 NLP (IDMC), Université de Lorraine, France.
  • Maxime Amblard and Amandine Decker, Dialogue Engineering, 14h, M2 NLP (IDMC) LI, Université de Lorraine, France.
  • Maxime Amblard and Marie Cousin, Math and Theoretical Computer Science, 50h M1 NLP (IDMC), Université de Lorraine, France.
  • Maxime Amblard, Introduction to NLP, M1 NLP (IDMC), Université de Lorraine, France.
  • Hee-Soo Choi, Written Corpora (English), 30h, M1 NLP (IDMC), Université de Lorraine, France.
  • Hee-Soo Choi, Software Projects (English), 24h, M1 NLP (IDMC), Université de Lorraine, France.
  • Fanny Ducel, Python Programming (English), 37h, M1 NLP (IDMC), Université de Lorraine, France.
  • Fanny Ducel, Introduction to G5K (English), 4h, M1 NLP (IDMC), Université de Lorraine, France.
  • Philippe de Groote and Marie Cousin Formal Logic, 22h, M1 NLP (IDMC), Université de Lorraine, France.
  • Philippe de Groote, Formal languages, 22h, M1 NLP (IDMC), Université de Lorraine, France.
  • Philippe de Groote and Amandine Decker, Semantics and Discourse, 30h, M2 NLP (IDMC), Université de Lorraine, France.
  • Philippe de Groote, Computational structures and logics for natural language modeling, 12h, MPRI, Université Paris Cité, France.
  • Karën Fort, Fanny Ducel, Ethics and NLP (English), 19h, M1 NLP (IDMC), Université de Lorraine, France.
  • Karën Fort, Fanny Ducel, Ethics and orientation (English), 25h, M2 NLP (IDMC), Université de Lorraine, France.
  • Karën Fort, Corpora, resources and tools for linguistics, 39h, M1, Sorbonne Université, France.
  • Bruno Guillaume, Written Corpora (English), 15h, M1 NLP (IDMC), Université de Lorraine, France.
  • Bruno Guillaume, Lexical Resources (English), 15h, M2 NLP (IDMC), Université de Lorraine, France.
  • Vincent Martin, Voice biomarkers (English), 12h, M2 NLP (IDMC), Université de Lorraine, France
  • Vincent Martin, NLP projects (English), 3h, M1 NLP (IDMC), Université de Lorraine, France
  • Vincent Martin, Critical analysis of artificial intelligence for health (English), 6h, Master 2 Health Engineering, Université Grenoble Alpes, France
  • Vincent Martin, Back to the big wide world: how to integrate digital tools into clinical practice? (English), 6h, Master 2 Health Engineering, Université Grenoble Alpes, France
  • Sylvain Pogodalla, Semantics, 10h, M1 NLP (IDMC), Université de Lorraine, France
  • Sylvain Pogodalla and Amandine Decker, Syntactic Models, 20h, M2 NLP (IDMC), Université de Lorraine, France
• Doctorate:
  • Karën Fort, Annotation collaborative de corpus, 4h, Summer school of the GDR LIFT, AnnoDemo 06 2024 – Banyuls-sur-Mer, France.
• International Summer School:
  • Bruno Guillaume with Daniel Zeman and Agata Savary, Corpus annotation infrastructure at the 1st UniDive training school, 8-12 July 2024 in Chișinău, Moldova.

10.2.2 Supervision

10.2.3 Juries

• Karën Fort PhD examiner and jury president for Célina Treuillier, Modelisation individuelle et multi-factorielle du phenomene de polarisation pour une personnalisation de l’apport en diversite dans les recommandations de news, Université de Lorraine, Oct. 2024.
• Karën Fort PhD examiner and jury president for Vijini Liyanage, Detection of artificially generated academic texts, Université Paris 13, May 2024.
• Karën Fort PhD reviewer (rapporteure) for Lilia Segundo Diaz, Juegos del español – A Collaborative Game-Based Approach to Building a Parsed Corpus of European Spanish Dialects, Hasselt University, Belgium, April 2024.
• Maxime Amblard PhD reviewer and jury member for Alban Petit, Structured prediction methods for semantic parsing, université Paris Saclay, France, February 2024
• Maxime Amblard PHD reviewer and president for Navneet Agarwal, Autοmated depressiοn level estimatiοn : a study οn discοurse structure, input representatiοn and clinical reliability, Université Normandie, June 2024
• Maxime Amblard PhD examiner and jury president for Esteban Marquer, Reasoning over Data : Analogy-based and Transfer Learning to improve Machine Learning, Université de Lorraine, France June 2024
• Maxime Amblard PHD examiner and jury president for Florian Marchal-Bornet, Towards Explaining Recommendation with Stories inside an Augmented Territory, Université de Lorraine, France, September 2024
• Michel Musiol PHD examiner and jury president for Mimault, T. Nouvelle perspective clinique des hommes auteurs de violences conjugales : Vers une typologie dissociative visant à mieux orienter la prévention de la récidive, joint supervision Université de Lorraine and Université du Québec à Trois-Rivières. November 2024

10.3.1 Productions (articles, videos, podcasts, serious games, ...)

• Karën Fort, Interview in Inria.fr: Pour une éthique du traitement automatique des langues : le regard de Karën Fort
• Karën Fort, Revue Vie de la recherche scientifique (VRS), 437, Les vrais dangers de l'IA
• Karën Fort, Maxime Amblard, Marc Anderson organised a doctoral training which allowed to creation of 6 short stories in 3 languages, forming a book 50
• Amandine Lecomte interview in France 3 Lorraine, April 2024 and France Info, April 2024, ”Santé : Comment l'intelligence artificielle peut aider les médecins.”

10.3.2 Participation in Live events

• Hee-Soo Choi: 2024-06-27, introducing linguistics and NLP to primary school children who have won the regional competition La Nouvelle de la Classe (ATILF, Nancy, France)
• Hee-Soo Choi and Fanny Ducel: 2024-01-16, presentation to high school students in the context of the "Chiche !" initiative (lycée Saint-Exupéry Fameck, France)
• Marie Cousin, Amandine Decker and Vincent Tourneur: 2024-01-11, presentation to high school students in the context of the "Chiche !" initiative (Lycée des métiers du tertiaire Jean-Victor Poncelet Saint-Avold, France)
• Marie Cousin, Amandine Decker, Karën Fort: 2024-09-21 and 2024-09-22, participation in “Journées européennes du Matrimoine” (Féru des Sciences, Nancy, France)
• Karën Fort, Invited presentation to the Association des Professeurs de Mathématiques de l'Enseignement Public (APMEP) : "Les outils de TAL : une opportunité pédagogique peut en cacher une autre", March 2024.
• Karën Fort, Apéro Scientifique du club ORION Human Interact - 10 avril 2024, Les enjeux éthiques de l'IA vus par le prisme du traitement automatique des langues
• Karën Fort, Mines students, LORIA - 17 janvier 2024, Enjeux éthiques du Traitement Automatique des Langues (TAL) : le cas des biais stéréotypés
• Karën Fort, IHEST, LORIA - 26 juin 2024, Enjeux éthiques de l'IA par le prisme du TAL
• Marie Cousin: 2024-02-02, participation in FIRST (Femmes Ingénieures, Réussir en Sciences et Technologies), présentation de la recherche à des élèves (filles) de seconde (Lycée Fabert, Metz, France)
• Valentin Richard gave a talk for Université Populaire et Participative de Vandœuvre at the Fabrique des Possibles (Vandœuvre-lès-Nancy), 17 April 2024.
• Maxime Amblard did a talk about AI for University libraries at Epinal
• Maxime Amblard did a talk about “Artificial Intelligence for schizophrenia » for high school students to prepare a presentation to the Comité Consultatif National d’Ethique, 22 mai 2024, Assemblée Nationale.
• Maxime Amblard did a talk ”IA et éthique : connaître les biais pour fixer les règles” for the national DPO day organised by CNIL and Metropole Grand Nancy, June 12th, 2024.
• Maxime Amblard presented Artificial Intellignece at la Frugalité Heureuse, national architecture meeting, October 6th, 2024.
• Maxime Amblard presented Artificial Intelligence for hight school students at Lycée Loritz, February 8th 2024.

International journals

• 10 articleM. M.Marc M Anderson and K.Karën Fort. Evaluating the acceptability of ethical recommendations in industry 4.0: an ethics by design approach. AI & Society: Knowledge, Culture and CommunicationJanuary 2024HALDOIback to text
• 11 articleF.Fanny Ducel, A.Aurélie Névéol and K.Karën Fort. "You'll be a nurse, my son!" Automatically Assessing Gender Biases in Autoregressive Language Models in French and Italian.Language Resources and EvaluationOctober 2024HALDOIback to text

National journals

• 12 articleF.Fanny Ducel, A.Aurélie Névéol and K.Karën Fort. Bias Research for Language Models is Biased: a Survey for Deconstructing Bias in Large Language Models.Revue TAL : traitement automatique des langues643September 2024, 119-143HALback to text
• 13 articleJ.-A.Jean-Arthur Micoulaud-Franchi, V.Vincent P. Martin, F.François Vialatte, C.Christophe Gauld, E.Elodie Gratreau and C.Clélia Quiles. La domestication de la sémiologie : proposition d’une organisation graphique du thesaurus semeioticus psychiatrique chez l’adulte.Annales Médico-Psychologiques, Revue PsychiatriqueNovember 2024HALDOIback to text

Invited conferences

• 14 inproceedingsH.-S.Hee-Soo Choi. French Lexical Semantic Graphs: Enrichment by Link Prediction and Integration in a WSD model.Séminaire annuel ANR SELEXINIParis, FranceJune 2024HALback to text

International peer-reviewed conferences

• 15 inproceedingsH.Hugo Boulanger, N.Nicolas Hiebel, O.Olivier Ferret, K.Karën Fort and A.Aurélie Névéol. Using Structured Health Information for Controlled Generation of Clinical Cases in French.The 6th Clinical Natural Language Processing Workshop At NAACL 2024 (ClinicalNLP 2024)Mexico city, MexicoJune 2024HALback to text
• 16 inproceedings H.-S.Hee-Soo Choi, P.Priyansh Trivedi, M.Mathieu Constant, K.Karën Fort and B.Bruno Guillaume. Beyond Model Performance: Can Link Prediction Enrich French Lexical Graphs? The 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING) Turin, Italy May 2024 HAL back to text
• 17 inproceedingsA.Amandine Decker and M.Maxime Amblard. With a Little Help from my (Linguistic) Friends: Topic Segmentation of Multi-party Casual Conversations.Proceedings of the 5th Workshop on Computational Approaches to Discourse (CODI 2024)Malta, MaltaAssociation for Computational LinguisticsMarch 2024, 177--188HALback to textback to text
• 18 inproceedingsA.Amandine Decker, V.Vincent Tourneur, M.Maxime Amblard and E.Ellen Breitholtz. "Wait, did you mean the doctor?": Collecting a Dialogue Corpus for Topical Analysis.Proceedings of the 28th Workshop on the Semantics and Pragmatics of DialogueSEMDIAL 2024Rovereto, Italy2024HALback to textback to text
• 19 inproceedingsK.Karën Fort, L. A.Laura Alonso Alemany, L.Luciana Benotti, J.Julien Bezançon, C.Claudia Borg, M.Marthese Borg, Y.Yongjian Chen, F.Fanny Ducel, Y.Yoann Dupont, G.Guido Ivetta, Z.Zhijian Li, M.Margot Mieskes, M.Marco Naguib, Y.Yuyan Qian, M.Matteo Radaelli, W. S.Wolfgang S. Schmeisser-Nieto, E.Emma Raimundo Schulz, T.Thiziri Saci, S.Sarah Saidi, J.Javier Torroba Marchante, S.Shilin Xie, S. E.Sergio E. Zanotto and A.Aurélie Névéol. Your Stereotypical Mileage may Vary: Practical Challenges of Evaluating Biases in Multiple Languages and Cultural Contexts.The 2024 Joint International Conference on Computational Linguistics, Language Resources and EvaluationTurin (Italie), ItalyMay 2024HALback to text
• 20 inproceedingsB.Bruno Guillaume, K.Kim Gerdes, K.Kirian Guiller, S.Sylvain Kahane and Y.Yixuan Li. Joint Annotation of Morphology and Syntax in Dependency Treebanks.The 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING)Turino, ItalyMay 2024HALback to text
• 21 inproceedingsM.Maxime Guillaume, S.Sylvain Pogodalla and V.Vincent Tourneur. ACGtk: A Toolkit for Developing and Running Abstract Categorial Grammars.Functional and Logic Programming. 17th International Symposium, FLOPS 2024, ProceedingsLecture Notes in Computer ScienceFunctional and Logic Programming. 17th International Symposium, FLOPS 2024LNCS 14659Kumamoto, JapanSpringerMay 2024, 13-30HALDOIback to text
• 22 inproceedingsN.Nicolas Hiebel, B.Bertrand Remy, B.Bruno Guillaume, O.Olivier Ferret, A.Aurélie Névéol and K.Karën Fort. Hostomytho: A GWAP for Synthetic Clinical Texts Evaluation and Annotation.Games and Natural Language Processing Workshop at LREC-COLING 2024, May 2024, Turin, ItalyTurin (Italie), ItalyMay 2024HALback to text
• 23 inproceedingsJ.Jacques Jayez. Argumentation et probabilités, ou pourquoi l'argumentation rationnelle n'est pas (toujours) un raisonnement.Congrès Mondial de Linguistique FrançaiseLausanne, FranceJuly 2024HALback to text
• 24 inproceedingsC.Chuyuan Li, C.Chloé Braud, M.Maxime Amblard and G.Giuseppe Carenini. Discourse Relation Prediction and Discourse Parsing in Dialogues with Minimal Supervision.Proceedings of the 5th Workshop on Computational Approaches to Discourse (CODI 2024)Malte, MaltaAssociation for Computational LinguisticsMarch 2024, 161--176HALback to text
• 25 inproceedingsA. C.Abelardo Carlos Martínez Lorenzo, P.-L.Pere-Luís Huguet Cabot, K.Karim Ghonim, L.Lu Xu, H.-S.Hee-Soo Choi, A. F.Alberte Fernández Castro and R.Roberto Navigli. Mitigating Data Scarcity in Semantic Parsing across Languages: the Multilingual Semantic Layer and its Dataset.The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024)The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024)Bangkok, ThailandAugust 2024HALback to text
• 26 inproceedingsA.Alice Millour, Y.Yoann Dupont, K.Karën Fort and L.Liam Duignan. Unveiling Strengths and Weaknesses of NLP Systems Based on a Rich Evaluation Corpus: the Case of NER in French.Proceedings of LREC-COLING 2024LREC-COLING 2024Turin, ItalyMay 2024HALback to text
• 27 inproceedingsS.Siyana Pavlova, M.Maxime Amblard and B.Bruno Guillaume. YARN is All You Knit: Encoding Multiple Semantic Phenomena with Layers.Proceedings of The Fifth International Workshop in Designing Meaning RepresentationThe Fifth International Workshop in Designing Meaning RepresentationTurin, ItalyMay 2024HALback to text
• 28 inproceedingsV. D.Valentin D. Richard. "selon comment vous vous positionnez" : Study of French Interrogative-based Adverbial Adjuncts.Actes du 9e Congrès Mondial de Linguistique Française9e Congrès Mondial de Linguistique Française191Lausanne, SwitzerlandSHS Web of ConferencesJune 2024, 14010HALDOIback to text
• 29 inproceedingsV. D.Valentin D. Richard. Dynamic Effects of Modalized Questions.Proceedings of the 24th Amsterdam ColloquiumProceedings of the 24th Amsterdam Colloquium298-307Amsterdam, NetherlandsDecember 2024HALback to text
• 30 inproceedingsV. D.Valentin D. Richard, E.Emmanuel Fruchard and V.Valentin Gatien-Baron. COORTE: A Toolkit for Putting a French Spelling Reform Suggestion into Practice.Actes du 9e Congrès Mondial de Linguistique Française9e Congrès Mondial de Linguistique Française191Lausanne, SwitzerlandSHS Web of ConferencesJune 2024, 11002HALDOI
• 31 inproceedingsA.Agata Savary, D.Daniel Zeman, V.Verginica Barbu Mititelu, A.Anabela Barreiro, O.Olesea Caftanatov, M.-C. D.Marie-Catherine De Marneffe, K.Kaja Dobrovoljc, G.Gülşen Eryiğit, V.Voula Giouli, B.Bruno Guillaume, S.Stella Markantonatou, N.Nurit Melnik, J.Joakim Nivre, A. K.Atul Kr. Ojha, C.Carlos Ramisch, A.Abigail Walsh, B.Beata Wójtowicz and A.Alina Wróblewska. UniDive: A COST Action on Universality, Diversity and Idiosyncrasy in Language Technology.Proceedings of the 3rd Annual Meeting of the Special Interest Group on Under-resourced Languages (SIGUL-2024) @ LREC-COLING 20243rd Annual Meeting of the Special Interest Group on Under-resourced LanguagesTorino, Italy2024HALback to text
• 32 inproceedingsE.Emmett Strickland, A.Anne Lacheret-Dujour, S.Sylvain Kahane, M.Marc Evrard, P.Perrine Quennehen, B.Bernard Caron, F.Francis Egbokhare and B.Bruno Guillaume. New Methods for Exploring Intonosyntax: Introducing an Intonosyntactic Treebank for Nigerian Pidgin.Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)Turin, ItalyACLMay 2024, 12207–12216HALback to text

National peer-reviewed Conferences

• 33 inproceedingsH.Hugo Boulanger, N.Nicolas Hiebel, O.Olivier Ferret, K.Karën Fort and A.Aurélie Névéol. Génération contrôlée de cas cliniques en français à partir de données médicales structurées.Actes de JEP-TALN-RECITAL 2024. Actes des 35èmes Journées d'Études sur la Parole35èmes Journées d'Études sur la Parole (JEP 2024) 31ème Conférence sur le Traitement Automatique des Langues Naturelles (TALN 2024) 26ème Rencontre des Étudiants Chercheurs en Informatique pour le Traitement Automatique des Langues (RECITAL 2024)Toulouse, FranceATALA & AFPCJuly 2024, 435-448HALback to text
• 34 inproceedingsA.Aaron Boussidan, F.Fanny Ducel, A.Aurélie Névéol and K.Karën Fort. What ChatGPT tells us about ourselves.Journée d'étude Éthique et TAL 2024Nancy, FranceApril 2024HALback to text
• 35 inproceedingsH.-S.Hee-Soo Choi, P.Priyansh Trivedi, M.Mathieu Constant, K.Karën Fort and B.Bruno Guillaume. Au-delà de la performance des modèles : la prédiction de liens peut-elle enrichir des graphes lexico-sémantiques du français ?Actes de JEP-TALN-RECITAL 2024. 31ème Conférence sur le Traitement Automatique des Langues Naturelles, volume 1 : articles longs et prises de positionToulouse, FranceATALA & AFPC2024, 36-49HALback to text
• 36 inproceedingsA.Amandine Decker and M.Maxime Amblard. Building an Unsupervised Topical Similarity Measure for Conversation.35èmes Journées d'Études sur la Parole (JEP 2024)35èmes Journées d'Études sur la Parole (JEP 2024) 31ème Conférence sur le Traitement Automatique des Langues Naturelles (TALN 2024) 26ème Rencontre des Étudiants Chercheurs en Informatique pour le Traitement Automatique des Langues (RECITAL 2024)1 : articles longs et prises de positionToulouse, FranceATALA & AFPC2024, 362-375HALback to text
• 37 inproceedingsF.Fanny Ducel, A.Aurélie Névéol and K.Karën Fort. Automatically Assessing Gender Biases in Autoregressive Language Models..TALN 2024Toulouse, FranceJuly 2024HALback to text
• 38 inproceedingsN.Nicolas Hiebel, O.Olivier Ferret, K.Karën Fort and A.Aurélie Névéol. Où la frugalité rejoint l'éthique : utilisation de données synthétiques pour la reconnaissance d'entités cliniques.Journée d'étude sur le traitement automatique des langues frugal et la recherche d'information frugale, ATALAParis, FranceJanuary 2024HALback to text
• 39 inproceedingsE.Emmett Strickland, A.Anne Lacheret-Dujour, M.Marc Evrard, S.Sylvain Kahane, D.Dana Aubakirova, D.Dorin Doncenco, D.Diego Torres, P.Perrine Quennehen and B.Bruno Guillaume. De nouvelles méthodes pour l'exploration de l'interface syntaxe-prosodie : un treebank intonosyntaxique et un système de synthèse pour le pidgin nigérian.35èmes Journées d'Études sur la Parole (JEP 2024)35èmes Journées d'Études sur la Parole (JEP 2024) 31ème Conférence sur le Traitement Automatique des Langues Naturelles (TALN 2024) 26ème Rencontre des Étudiants Chercheurs en Informatique pour le Traitement Automatique des Langues (RECITAL 2024)1 : articles longs et prises de positionToulouse, FranceATALA & AFPC2024, 376-383HALback to text

Conferences without proceedings

• 40 inproceedingsC.Clémentine Bleuze, F.Fanny Ducel, K.Karën Fort and M.Maxime Amblard. Vers la création d'une super-intelligence » : un corpus pour étudier les revendications des articles de TAL.Journées de lancement LIFT 2Orléans, FranceNovember 2024HALback to text
• 41 inproceedingsF.Fanny Ducel, A.Aurélie Névéol and K.Karën Fort. Desiderata for Actionable Bias Research.New Perspectives on Bias and Discrimination in Language TechnologyAmsterdam (Pays-Bas), FranceNovember 2024HALback to text
• 42 inproceedingsP.Philippe de Groote and T.Timothée Bernard. Perspective on individuals.Sinn und Bedeutung 29Noto, Italy2024HALback to text
• 43 inproceedingsA.Althea Löfgren, S.Santiago Herrera, S.Sylvain Kahane, B.Bruno Guillaume and N.Natalia Levshina. Exploring Sampling Strategies for Linguistic Diversity: A Comparative Analysis of UD Treebanks.15th International Conference of the Association for Linguistic TypologySingapore (SG), SingaporeDecember 2024HALback to text

Scientific book chapters

• 44 inbookP.Philippe de Groote, M.Maxime Guillaume, A.Agathe Helman, S.Sylvain Pogodalla and R.Raphaël Salmon. Extending Abstract Categorial Grammars with Feature Structures: Theory and Practice.14569Logic and Engineering of Natural Language Semantics, 20th International Conference, LENLS20, Osaka, Japan, November 18–20, 2023, Revised Selected PapersLecture Notes in Computer ScienceSpringer2024, 118-133HALDOIback to text
• 45 inbookP.Philippe de Groote. On the Semantics of Dependencies: Relative Clauses and Open Clausal Complements.14569Logic and Engineering of Natural Language Semantics, 20th International Conference, LENLS20, Osaka, Japan, November 18–20, 2023, Revised Selected PapersLecture Notes in Computer ScienceSpringer2024, 244-259HALDOIback to text

Edition (books, proceedings, special issue of a journal)

• 46 proceedingsEthics and NLP: 10 years after.Journée d'études ATALA "éthique et TAL : 10 ans après"2024HALback to text

Reports & preprints

• 47 miscM.Marie Cousin. Adding communicative structure to the MMT into ACG encoding.October 2024HALback to text

Scientific popularization

• 48 inbookH.-S.Hee-Soo Choi, M.Marie Cousin, A.Amandine Decker and V. D.Valentin D Richard. PENSE M'EN.THINK BEFORE LOADING2024HAL
• 49 inbookG.Guillaume Coiffier, S.Sewade Ogun, L.Leo Valque and P.Priyansh Trivedi. STATE OF THE ART.THINK BEFORE LOADING2024HAL
• 50 bookK.Karën Fort, M. M.Marc M Anderson, A.Aurore Coince, M.Mathieu D’aquin, M.Maxime Amblard, S. E.Sarah E. Carter, I.Ilaria Tiddi and D.Diane Ranville, eds. Think before loading: Ne vous en faites pas, ça va mal se passer.2024HALback to textback to text