2025Activity​‌ reportProject-TeamALMANACH

Node infrastructure:

We​​ have access to three​​​‌ types of GPU node​ on Jean Zay:12​‌

1. Nodes comprising 4 Nvidia​​ Tesla V100 SXM2 32GB​​​‌ GPUs, 192GB of RAM,​ and two Intel Cascade​‌ Lake 6248 processors. One​​ Nvidia Tesla V100 card​​​‌ is rated at around​ 300W,13 while the​‌ Intel Cascade Lake processor​​ is rated at 150W.​​​‌14 For the DRAM​ we can use the​‌ work of 101 to​​ estimate the total power​​​‌ draw of 192GB of​ RAM at approximately 20W.​‌ The total power draw​​ of one Jean Zay​​​‌ node at peak use​ therefore adds up to​‌ around 1520W.
2. Nodes comprising​​ 8 Nvidia A100 SXM4​​​‌ 80GB GPUs, 512GB of​ RAM, and two AMD​‌ Milan EPYC 7543 processors.​​ One Nvidia A100 card​​​‌ is rated at around​ 400W.15 while the​‌ AMD Milan processor is​​ rated at 225W.16​​​‌ Following 101, we​ estimate the total power​‌ draw of 512GB of​​ RAM at approximately 50W.​​​‌ The total power draw​ of one A100 node​‌ at peak use therefore​​ adds up to around​​​‌ 3700W.
3. Nodes comprising 4​ Nvidia H100 SXM5 80GB​‌ GPUs, 512GB of RAM​​ and two Intel Xeon​​​‌ Platinum 8468 processors. One​ Nvidia H100 card is​‌ rated around 700W 17​​, while the Intel​​​‌ Xeon Platinum 8468 processor​ is rated at 300W.​‌18 The total power​​ draw of one H100​​​‌ node at peak use​ therefore adds up to​‌ around 3450W.

We have​​ recently gained access to​​​‌ the Adastra supercomputer at​ CINES,19 which has​‌

1. nodes comprising 4 AMD​​ MI250Xs and 1 AMD​​​‌ EPYC Trento processor. Adastra​ uses the same hardware​‌ as the LUMI cluster,​​20 for which the​​​‌ total power draw of​ one node at peak​‌ use has been reported​​ as being approximately 2500W.​​​‌21
2. nodes comprising 4​ AMD MI300As (APUs including​‌ integrated AMD Zen 4​​ x86 CPU cores), for​​​‌ which the total power​ draw of one at​‌ node at peak has​​ been reported as being​​​‌ approximately 3,125W22

With​ this information, we use​‌ the formula proposed by​​ 124 and compute the​​​‌ total power required for​ each setting:

Where​​ $c$ and $g$ are​​​‌ the number of CPUs​ and GPUs respectively, ${\mathrm{p‌}}_c$ is the average​​ power draw (in W)​​​‌ from all CPU sockets,​ $p_r$ the average​‌ power draw from all​​ DRAM sockets, and ${\mathrm{p‌}}_g$ the average power​ draw of a single​‌ GPU. We estimate the​​ total power consumption by​​​‌ adding GPU, CPU and​ DRAM consumption, and then​‌ multiplying by the Power​​ Usage Effectiveness (PUE), which​​ accounts for the additional​​​‌ energy required to support‌ the compute infrastructure. We‌​‌ use a PUE coefficient​​ of 1.20, which is​​​‌ the value reported by‌ IDRIS for the Jean‌​‌ Zay supercomputer. For the​​ real time $t$ we​​​‌ have to divide the‌ reported time for each‌​‌ Jean Zay project by​​ 4 for V100 and​​​‌ H100 nodes and 8‌ for A100 nodes, as‌​‌ Jean Zay reports the​​ computing time of each​​​‌ project in GPU hours‌ and not in per-node‌​‌ hours. In Table 1​​ we report the training​​​‌ times in hours, as‌ well as the total‌​‌ power draw (in kWh)​​ of each Jean Zay​​​‌ and Adastra project associated‌ with the ALMAnaCH team‌​‌ during 2025. We use​​ this information to compute​​​‌ the total power consumption‌ (multiplying by the PUE)‌​‌ of each project, also​​ reported in Table 1​​​‌.

We can further‌ estimate the CO${}_{\mathrm{2‌‌}}$ emissions in kilograms of​​ each single project by​​​‌ multiplying the total power‌ consumption by the average‌​‌ CO${}_2$ emissions per​​ kWh in France, which​​​‌ were around 27g/kWh on‌ average for 2025.23‌​‌ The total CO${}_{\mathrm{2}}$ emissions in kg for​​​‌ one single model can‌ therefore be computed as:‌​‌

All emissions‌ are also reported in‌​‌ Table 1. The​​ total emission estimate for​​​‌ the team is 20,106kg‌ of CO${}_2$.‌​‌ The carbon footprint of​​ a single passenger on​​​‌ a round trip from‌ Paris to Albuquerque (USA)‌​‌ (Boeing 787), flying economy,​​ amounts to around 3,400kg​​​‌ of CO${}_2$.‌24 This means that‌​‌ our computing emissions from​​ Jean Zay and Adastra​​​‌ for 2025 amount to‌ under six Paris-Albuquerque trips‌​‌ (NAACL 2025), and therefore,​​ given the size of​​​‌ our project-team, the largest‌ source of emissions from‌​‌ the team are from​​ flights to conferences, which​​​‌ are a necessary part‌ of communicating about our‌​‌ research to the international​​ community.

7.1.1​​​‌ HTR-United

• Keywords:
  HTR, OCR​
• Functional Description:

  HTR-United is​‌ a Github organization without​​ any other form of​​​‌ legal personality. It aims​ at gathering HTR/OCR transcriptions​‌ of all periods and​​ styles of writing, mostly​​​‌ but not exclusively in​ French. It was born​‌ from the mere necessity​​ for projects- to possess​​​‌ potentiel ground truth to​ rapidly train models on​‌ smaller corpora.

  Datasets shared​​ or referenced with HTR-United​​​‌ must, at minimum, take​ the form of: (i)​‌ an ensemble of ALTO​​ XML and/or PAGE XML​​​‌ files containing either only​ information on the segmentation,​‌ either the segmentation and​​ the corresponding transcription, (ii)​​​‌ an ensemble of corresponding​ images. They can be​‌ shared in the form​​ of a simple permalink​​​‌ to ressources hosted somewhere​ else, or can be​‌ the contact information necessary​​ to request access to​​​‌ the images. It must​ be possible to recompose​‌ the link between the​​ XML files and the​​​‌ image without any intermediary​ process, (iii) a documentation​‌ on the transcription practices​​ followed for the segmentation​​​‌ and the transcription. In​ the cases of a​‌ Github repository, this documentation​​ must be summarized in​​​‌ the README.

  A corpus​ can be sub-diveded into​‌ smaller ensembles if it​​ seems necessary.

• Release Contributions:​​​‌
  First version.
• URL:
  https://­htr-united.­github.­io/​
• Contact:
  Alix Chague

7.1.2​‌ CamemBERT-bio

• Keywords:
  Language model,​​ Deep learning, NLP, Transformer​​​‌
• Functional Description:
  CamemBERT-bio is​ a state-of-the-art french biomedical​‌ language model built using​​ continual-pretraining from camembert-base. It​​​‌ was trained on a​ french public biomedical corpus​‌ of 413M words containing​​ scientific documments, drug leaflets​​​‌ and clinical cases extrated​ from theses and articles.​‌ It shows significant improvement​​ on multiple biomedical named​​​‌ entity recognition tasks compared​ to camembert-base.
• URL:
  http://­camembert-bio-model.­fr/​‌
• Contact:
  Rian Touchent

7.1.3​​ RoCS-MT

• Name:
  Robust Challenge​​​‌ Set for Machine Translation​
• Keywords:
  Machine translation, NLP,​‌ Evaluation, Robustness, User-generated content​​
• Functional Description:
  RoCS-MT, a​​​‌ Robust Challenge Set for​ Machine Translation (MT), is​‌ designed to test MT​​ systems’ ability to translate​​​‌ user-generated content (UGC) that​ displays non-standard characteristics, such​‌ as spelling errors, devowelling,​​ acronymisation, etc. RoCS-MT is​​​‌ composed of English comments​ from Reddit, selected for​‌ their non-standard nature, which​​ have been manually normalised​​​‌ and professionally translated into​ five languages: French, German,​‌ Czech, Ukrainian and Russian.​​ The challenge set was​​​‌ included as a test​ suite at the WMT​‌ 2023 conference. This repository​​ therefore also includes automatic​​​‌ translations from the submissions​ to the general MT​‌ task.
• URL:
  https://­github.­com/­rbawden/­RoCS-MT
• Publications:​​
  hal-04300824, hal-05344725
• Contact:​​​‌
  Rachel Bawden
• Participants:
  Rachel​ Bawden, Benoit Sagot

7.1.4​‌ 3MT_French Dataset

• Name:
  3​​ Minutes Thesis Corpus
• Keywords:​​​‌
  Multimodal Corpus, Video annotation​
• Functional Description:
  This new​‌ resource will be useful​​ to computer science and​​​‌ social science researchers working​ on public speaking assessment​‌ and training. It will​​ help refine the analysis​​​‌ of speaking from a​ fresh perspective based on​‌ social-cognitive theories rarely studied​​ in this context, such​​​‌ as first impressions and​ theories of primacy and​‌ recency.
• URL:
  https://­zenodo.­org/­records/­7603511#.­Y904l3CZOUk
• Publication:​​
  hal-04366763
• Contact:
  Chloe Clavel​​​‌

7.1.5 CATMuS Medieval (Model)​

• Name:
  Consistent Approach to​‌ Transcribing ManuScripts - Medieval​​ model
• Keyword:
  Handwritten Text​​ Recognition
• Functional Description:
  CATMuS​​​‌ (Consistent Approach to Transcribing‌ ManuScript) Medieval is a‌​‌ model for automatically transcribing​​ medieval manuscripts using Latin​​​‌ scripts, in particular Old‌ and Middle French, Latin,‌​‌ Spanish (and other languages​​ of Spain), and Italian.​​​‌ The model was trained‌ on the largest and‌​‌ most diverse dataset known​​ for medieval manuscripts in​​​‌ Latin scripts, with more‌ than 110 000 lines‌​‌ of training data.
• Contact:​​
  Thibault Clerice
• Partners:
  University​​​‌ of Toronto, Ecole nationale‌ des chartes, CIHAM UMR‌​‌ 5648, VeDPH - Ca'​​ Foscari, Université de Genève,​​​‌ ENS Lyon

7.1.6 HTRomance‌

• Keyword:
  Handwritten Text Recognition‌​‌
• Functional Description:
  The ground​​ truth produced as part​​​‌ of the HTRomance project‌ aims to provide diverse‌​‌ data, from the 12th​​ century to the 19th​​​‌ century, for training handwritten‌ text recognition models. It‌​‌ covers the following languages:​​ Latin, various states of​​​‌ French, Spanish, Occitan and‌ Italian.
• URL:
  https://­htromance-project.­github.­io/
• Contact:‌​‌
  Thibault Clerice
• Partners:
  VeDPH​​ - Ca' Foscari, Ecole​​​‌ nationale des chartes, CIHAM‌ UMR 5648, ENS Lyon‌​‌

7.1.7 eScriptorium Documentation

• Name:​​
  Open documentation for eScriptorium​​​‌
• Functional Description:
  Collaborative and‌ open documentation redacted using‌​‌ the functionalities offered by​​ Github et deployed thanks​​​‌ to Readthedocs. It offers‌ an illustrated description of‌​‌ all the features of​​ the eScriptorium application, which​​​‌ does not offer any‌ other form of complete‌​‌ documentation, as well as​​ tutorials.
• URL:
  https://­escriptorium.­readthedocs.­io/
• Contact:​​​‌
  Alix Chague

7.1.8 CATMuS‌ Medieval (Dataset)

• Name:
  Consistent‌​‌ Approaches to Transcribing ManuScripts​​ - Medieval Dataset
• Keywords:​​​‌
  Handwritten Text Recognition, HTR,‌ OCR
• Functional Description:
  Developed‌​‌ through collaboration among various​​ institutions and projects, CATMuS​​​‌ provides an inter-compatible handwritten‌ text recognition dataset spanning‌​‌ more than 240 manuscripts​​ and incunabula in 10​​​‌ different languages, comprising over‌ 170,000 lines of text‌​‌ and 5 million characters​​ spanning from the 8th​​​‌ century to the 16th.‌
• Release Contributions:
  - 40‌​‌ new mansucripts
• Publication:
  hal-04453952​​
• Contact:
  Thibault Clerice
• Partners:​​​‌
  CIHAM UMR 5648, Ecole‌ nationale des chartes, University‌​‌ of Toronto, Antwerp University​​

7.1.9 Counter dataset

• Name:​​​‌
  Counter Radicalization Dataset
• Keywords:‌
  NLP, Multilingual corpus, Data‌​‌ protection, Pseudonymization, Online radicalization​​
• Functional Description:
  This dataset​​​‌ includes multilingual content from‌ forums, Telegram, social media,‌​‌ and other sources in​​ English, French, and Arabic.​​​‌ It covers various radical‌ ideologies and is pseudonymized‌​‌ to protect privacy while​​ maintaining data utility. It​​​‌ includes annotations for call‌ to action, radicalization level,‌​‌ and named entity recognition.​​
• URL:
  https://­gitlab.­inria.­fr/­ariabi/­counter-dataset-public
• Contact:
  Djame​​​‌ Seddah

7.1.10 CamemBERTav2

• Keywords:‌
  Language model, French
• Functional‌​‌ Description:

  It is the​​ second version of the​​​‌ CamemBERTa model, which is‌ based on the DebertaV2‌​‌ architecture with the Replaced​​ Token Detection (RTD) objective.​​​‌

  The new update includes:‌ 1) Much larger pretraining‌​‌ dataset: 275B unique tokens​​ (previously 2B). 2) A​​​‌ newly built tokenizer based‌ on WordPiece with 32,768‌​‌ tokens, addition of the​​ newline and tab characters,​​​‌ support emojis, and better‌ handling of numbers (numbers‌​‌ are split into two​​ digits tokens). 3) Extended​​​‌ context window of 1024‌ tokens

• Contact:
  Benoit Sagot‌​‌

7.1.11 CamemBERTv2

• Keywords:
  Language​​ model, French
• Functional Description:​​​‌
  CamemBERTv2 is a French‌ language model pretrained on‌​‌ a large up-to-date corpus​​​‌ of 275B tokens of​ French text. The model​‌ still based on the​​ BERT architecture, demonstrates improved​​​‌ performance across a variety​ of NLP tasks, over​‌ the previous version
• URL:​​
  https://­huggingface.­co/­almanach/­camembertv2-base
• Contact:
  Benoit Sagot​​​‌

7.1.12 LADaS

• Name:
  Layout​ Analysis Dataset with Segmonto​‌
• Keyword:
  Layout Analysis
• Functional​​ Description:
  LADaS is a​​​‌ dataset for training layout​ analysis on documents from​‌ the 16th to the​​ 21st century. It helps​​​‌ reproduce documents in XML​ TEI.
• Publication:
  hal-04513725
• Contact:​‌
  Thibault Clerice

7.1.13 CamemBERT-bio-gliner​​

• Keywords:
  Language model, Deep​​​‌ learning, NLP, Transformer
• Functional​ Description:
  CamemBERT-bio-gliner is a​‌ Named Entity Recognition (NER)​​ model capable of identifying​​​‌ any french biomedical entity​ type using a BERT-like​‌ encoder. It provides a​​ practical alternative to traditional​​​‌ NER models, which are​ limited to predefined entities,​‌ and Large Language Models​​ (LLMs) that, despite their​​​‌ flexibility, are costly and​ large for resource-constrained scenarios.​‌ CamemBERT-bio is used as​​ a backbone.
• Contact:
  Rian​​​‌ Touchent

7.1.14 CoMMuTE

• Name:​
  Contrastive multilingual and multimodal​‌ translation evaluation
• Keywords:
  Machine​​ translation, Evaluation, Image analysis​​​‌
• Functional Description:
  CoMMuTE is​ a contrastive evaluation dataset​‌ designed to assess the​​ ability of multimodal machine​​​‌ translation models to exploit​ images in order to​‌ disambiguate the sentence to​​ be translated. In other​​​‌ words, given a sentence​ containing a word that​‌ can be translated in​​ several ways, the additional​​​‌ image determines the meaning​ of the word to​‌ be translated. The model​​ must then take the​​​‌ image into account to​ propose a correct translation.​‌ CoMMuTE is available from​​ English into French, German​​​‌ and Czech.
• URL:
  https://­github.­com/­MatthieuFP/­CoMMuTE​
• Publications:
  hal-03977982, hal-04736377​‌
• Contact:
  Matthieu Futeral-Peter
• Participants:​​
  Matthieu Futeral-Peter, Rachel Bawden,​​​‌ Benoit Sagot

7.1.15 mOSCAR​

• Keywords:
  Raw corpus, Multilingual​‌ corpus, Multimodal Corpus, Multimodality,​​ Text-image processing, Web crawling​​​‌
• Functional Description:
  mOSCAR is​ a the first large-scale​‌ multilingual and multimodal web-crawled​​ document corpus. It covers​​​‌ 163 languages, 315M documents,​ 214B tokens and 1.2B​‌ images. We carefully conduct​​ a set of filtering​​​‌ and evaluation steps to​ make sure mOSCAR is​‌ sufficiently safe, diverse and​​ of good quality.
• URL:​​​‌
  https://­oscar-project.­github.­io/­documentation/­versions/­mOSCAR/
• Publication:
  hal-04629451
• Contact:​
  Matthieu Futeral-Peter
• Participants:
  Matthieu​‌ Futeral-Peter, Benoit Sagot, Rachel​​ Bawden, Julien Abadji, Armel​​​‌ Zebaze Dongmo, Cordelia Schmid,​ Remi Lacroix

7.1.16 Concordancer​‌

• Name:
  Concordancer
• Keywords:
  Natural​​ language processing, Document analysis,​​​‌ Terminological analysis
• Scientific Description:​
  Concordancer enables the detection​‌ in a parsed corpus​​ of a set of​​​‌ terms from a terminology.​ It also detects potential​‌ variants of these terms​​ by applying various transformations​​​‌ (such as reduction, acronym,​ morphological derivations, synonymy, etc.),​‌ in order to study​​ variations in the use​​​‌ of a term within​ a same document.
• Functional​‌ Description:
  Concordancer enables the​​ detection in a parsed​​​‌ corpus of a set​ of terms from a​‌ terminology. It also detects​​ potential variants of these​​​‌ terms by applying various​ transformations (such as reduction,​‌ acronym, morphological derivations, synonymy,​​ etc.), in order to​​​‌ study variations in the​ use of a term​‌ within a same document.​​
• Contact:
  Éric De La​​​‌ Clergerie

7.1.17 SWELLS

• Name:​
  Specialized Workbench for the​‌ Explicit Learning of Linguistic​​ Structures
• Keywords:
  Machine translation,​​ Natural language processing, Low​​​‌ resources languages, Large Language‌ Models, Language workbench
• Functional‌​‌ Description:
  SWELLS is a​​ framework specially designed to​​​‌ study explicit learning in‌ Large Language Models (LLMs)‌​‌ by means of machine​​ translation experiments involving constructed​​​‌ languages (conlangs). It provides‌ datasets and scripts that‌​‌ enable a controlled assessment​​ of an LLM’s ability​​​‌ to learn and apply‌ metalinguistic rules presented in‌​‌ grammar books.
• URL:
  https://­github.­com/­mmarmonier/­SWELLS​​
• Publication:
  hal-04991098
• Contact:
  Malik​​​‌ Marmonier
• Participants:
  Malik Marmonier,‌ Rachel Bawden, Benoit Sagot‌​‌

7.1.18 ACReFOSC

• Name:
  A​​ Companion Repository to the​​​‌ French OLDI Seed Corpus‌
• Keyword:
  Machine translation
• Functional‌​‌ Description:
  ACReFOSC provides the​​ code and data needed​​​‌ to automatically generate preference‌ datasets for machine translation,‌​‌ supporting research in preference​​ optimization and the post-training​​​‌ of neural models and‌ LLMs for machine translation‌​‌ purposes.
• URL:
  https://­github.­com/­mmarmonier/­ACReFOSC
• Publication:​​
  hal-05375157
• Contact:
  Malik Marmonier​​​‌
• Participants:
  Malik Marmonier, Rachel‌ Bawden, Benoit Sagot

7.1.19‌​‌ CoMMA

• Name:
  Corpus of​​ Multilingual Medieval Archives
• Keywords:​​​‌
  Historical language, Raw corpus,‌ TEI, XML, HTR, OCR‌​‌
• Functional Description:
  CoMMA is​​ a large-scale corpus of​​​‌ medieval manuscripts produced through‌ automatic text recognition. The‌​‌ corpus contains around 3.3b​​ tokens drawn from more​​​‌ than 32,700 digitized manuscripts‌ in Latin and Old‌​‌ French.
• Contact:
  Thibault Clerice​​
• Partners:
  Université de Genève,​​​‌ CIHAM UMR 5648

7.1.20‌ NeWMe

• Name:
  The NeWMe‌​‌ Corpus (Negotiating Word Meaning)​​
• Keywords:
  Raw corpus, Natural​​​‌ language, Dialogue, Semantics, Discourse‌
• Functional Description:
  The NeWMe‌​‌ corpus contains 661 annotated​​ examples of Word Meaning​​​‌ Negotiation (WMN: sequences in‌ conversation where speakers discuss‌​‌ word meaning) as well​​ as related conversational phenomena.​​​‌ Annotations contain the type‌ of phenomenon as well‌​‌ as parts of the​​ WMN: a trigger (problematic​​​‌ word usage), indicator (signaling‌ a problem with word‌​‌ meaning), and the negotiation​​ (a meta-linguistic discussion of​​​‌ word meaning). NeWMe contains‌ oral conversations from Switchboard‌​‌ and the BNC as​​ well as written online​​​‌ debates from Reddit.
• Contact:‌
  Aina Gari Soler

7.1.21‌​‌ SPOT

• Keyword:
  Social networks​​
• Functional Description:
  SPOT (Stopping​​​‌ Points in Online Threads)‌ is the first annotated‌​‌ corpus operationalizing the sociological​​ concept of stopping point​​​‌ (i.e critical interventions) as‌ a reproducible task for‌​‌ computational study. The SPOT​​ dataset comprises approximately 43,000​​​‌ comments from French Facebook‌ discussions linked to URLs‌​‌ flagged as false information,​​ enriched with contextual metadata​​​‌ (parent post, parent comment,‌ page/group and source features)‌​‌
• Contact:
  Celia Nouri

7.1.22​​ eScriptorium

• Keywords:
  HTR, OCR,​​​‌ Web Application, Annotation tool‌
• Functional Description:
  Web application‌​‌ for manual, semi-automatic, and​​ automatic segmentation and transcription​​​‌ of printed or handwritten‌ text documents, with the‌​‌ possibility of training or​​ reusing transcription models.
• News​​​‌ of the Year:
  Thibault‌ Clérice joined the boards‌​‌ of directors for the​​ development of eScriptorium. Hassen​​​‌ Aguili became the main‌ maintainer of the software‌​‌ in december 2025.
• Contact:​​
  Thibault Clerice
• Participant:
  Hassen​​​‌ Aguili

7.1.23 ocDI

• Keywords:‌
  Supervised models, Variety identification,‌​‌ Natural language processing
• Functional​​ Description:

  A series of​​​‌ models for the identification‌ of the Occitan dialect‌​‌ of a text. The​​ models were trained on​​​‌ several dialect-labelled Occitan datasets‌ to predict a dialect‌​‌ out of 8 labels​​​‌ corresponding to the following​ Occitan varieties: Languedocian, Provençal,​‌ Limousine, Auvergnat, Vivaro-alpine, Niçard,​​ Gascon, and Aranese.

  Three​​​‌ variants of the model​ are available to correspond​‌ to various usages and​​ constraints: - ocDI-SVM: a​​​‌ linear classifier based on​ full words and n-grams​‌ of characters, that enables​​ easy interpretation of the​​​‌ model predictions, and obtained​ the best recall for​‌ Auvergnat, Gascon, Niçard and​​ Vivaro-alpine. - ocDI-FastText: a​​​‌ very fast classifier based​ on word and character​‌ n-gram static embeddings, which​​ obtained the best recall​​​‌ for Gascon and Limousine.​ - ocDI-BERT: a fine-tuned​‌ BERT model that obtained​​ the best overall performance​​​‌

  For data mining purposes,​ it is recommended to​‌ use the model with​​ the best recall on​​​‌ the target variety as​ our models tend to​‌ over-predict the most represented​​ classes (esp. Languedocian).

• Contact:​​​‌
  Oriane Nedey
• Participants:
  Oriane​ Nedey, Thibault Clerice, Rachel​‌ Bawden, Benoit Sagot

7.1.24​​ OcWikiDialects

• Keywords:
  Raw corpus,​​​‌ Low resources languages
• Functional​ Description:
  OcWikiDialects is a​‌ corpus of over 7k​​ articles from the Occitan​​​‌ Wikipedia, with rich metadata​ including dialect labels for​‌ 8 varieties of Occitan​​ (Provençal, Languedocian, Gascon, Auvergnat,​​​‌ Limousine, Vivaro-alpine, Niçard, Aranese)​ and 2 transitional varieties​‌ (Marchese and Aguianese). It​​ contains full articles in​​​‌ Markdown and WikiText formats,​ and segmentations into paragraphs​‌ and sentences, for a​​ total of approximately 4M​​​‌ tokens.
• Contact:
  Oriane Nedey​

7.1.25 GAPeron

• Name:
  GAPeron​‌
• Keywords:
  LLM, LLM Safety​​
• Functional Description:
  We release​​​‌ Gaperon, a fully open​ suite of French-English-coding language​‌ models designed to advance​​ transparency and reproducibility in​​​‌ large-scale model training. The​ Gaperon family includes 1.5B,​‌ 8B, and 24B parameter​​ models trained on 2-4​​​‌ trillion tokens, released with​ all elements of the​‌ training pipeline: French and​​ English datasets filtered with​​​‌ a neural quality classifier,​ an efficient data curation​‌ and training framework, and​​ hundreds of intermediate checkpoints.​​​‌ Through this work, we​ study how data filtering​‌ and contamination interact to​​ shape both benchmark and​​​‌ generative performance. We find​ that filtering for linguistic​‌ quality enhances text fluency​​ and coherence but yields​​​‌ subpar benchmark results, and​ that late deliberate contamination​‌ – continuing training on​​ data mixes that include​​​‌ test sets – recovers​ competitive scores while only​‌ reasonably harming generation quality.​​ We discuss how usual​​​‌ neural filtering can unintentionally​ amplify benchmark leakage. To​‌ support further research, we​​ also introduce harmless data​​​‌ poisoning during pretraining, providing​ a realistic testbed for​‌ safety studies. By openly​​ releasing all models, datasets,​​​‌ code, and checkpoints, Gaperon​ establishes a reproducible foundation​‌ for exploring the trade-offs​​ between data curation, evaluation,​​​‌ safety, and openness in​ multilingual language model development.​‌
• URL:
  https://­huggingface.­co/­collections/­almanach/­gaperon
• Publication:
  hal-05410121​​
• Contact:
  Djame Seddah
• Participants:​​​‌
  Nathan Godey, Wissam Antoun,​ Rian Touchent, Djame Seddah,​‌ Éric De La Clergerie,​​ Benoit Sagot, Rachel Bawden​​​‌

7.1.26 ModernCamemBERT

• Keyword:
  LLM​
• Functional Description:
  The goal​‌ of ModernCamemBERT was to​​ run a controlled study​​​‌ by pretraining ModernBERT on​ the same dataset as​‌ CamemBERTaV2, a DeBERTaV3 French​​ model, isolating the effect​​​‌ of model design. Our​ results show that the​‌ previous model generation remains​​ superior in sample efficiency​​ and overall benchmark performance,​​​‌ with ModernBERT’s primary advantage‌ being faster training and‌​‌ inference speed. However, the​​ new proposed model still​​​‌ provides meaningful architectural improvements‌ compared to earlier models‌​‌ such as the BERT​​ and RoBERTa CamemBERT/v2 model.​​​‌
• URL:
  https://­huggingface.­co/­collections/­almanach/­moderncamembert
• Contact:
  Wissam‌ Antoun

7.1.27 Glyphea

• Keywords:‌​‌
  Web Application, Cryptography, HTR​​
• Functional Description:
  Web application​​​‌ for transcribing encrypted handwritten‌ texts
• Contact:
  Cécile Pierrot‌​‌

8.1.1 Language​​ Models

Our research activity​​​‌ on language modelling in‌ 2025 has focused on‌​‌ understanding, improving, and demystifying​​​‌ large-scale transformer models along​ three complementary axes: architectural​‌ evaluation under controlled conditions,​​ openness and data-centric analysis​​​‌ in large language model​ (LLM) training, and efficiency​‌ at inference time, particularly​​ for long-context generation.

A​​​‌ first line of work​ addresses a growing challenge​‌ in the field: disentangling​​ the respective contributions of​​​‌ model architecture and pretraining​ data to downstream performance.​‌ Recent transformer variants often​​ report gains over previous​​​‌ models, but these claims​ are frequently confounded by​‌ undisclosed data or incomparable​​ training setups. To address​​​‌ this, in the context​ of Wissam Antoun 's​‌ PhD thesis supervised by​​ Djamé Seddah and Benoît​​​‌ Sagot , we conducted​ a controlled comparison between​‌ two recent encoder architectures​​ by pretraining ModernBERT on​​​‌ exactly the same data​ as CamemBERTaV2, a DeBERTaV3-based​‌ French model 24.​​ This study shows that,​​​‌ when data is held​ constant, the previous model​‌ generation remains superior in​​ both sample efficiency and​​​‌ overall benchmark performance. ModernBERT's​ main advantages lie instead​‌ in engineering-oriented improvements, such​​ as support for longer​​​‌ contexts and faster training​ and inference. Beyond this​‌ comparison, we also show​​ that higher-quality pretraining data​​​‌ mainly accelerates convergence without​ substantially improving final performance,​‌ suggesting a degree of​​ saturation in commonly used​​​‌ benchmarks. Overall, this work​ advocates for more rigorous​‌ evaluation protocols and highlights​​ that architectural novelty should​​​‌ be assessed independently from​ data scale and quality.​‌

A second major contribution​​ is our commitment to​​​‌ transparency and reproducibility in​ large-scale language model training,​‌ embodied in the release​​ of the Gaperon model​​​‌ suite 76, a​ collaborative effort primarily supervised​‌ by Djamé Seddah and​​ Benoît Sagot and involving​​​‌ Nathan Godey , Wissam​ Antoun , Rian Touchent​‌ , Éric Villemonte De​​ La Clergerie and Rachel​​​‌ Bawden . Gaperon consists​ of fully open French-English-code​‌ autoregressive models ranging from​​ 1.5B to 24B parameters,​​​‌ trained on trillions of​ tokens. Crucially, we release​‌ not only the final​​ models, but also the​​​‌ datasets, filtering tools, training​ code, and hundreds of​‌ intermediate checkpoints. This openness​​ enables a systematic investigation​​​‌ of how data filtering,​ contamination, and evaluation practices​‌ interact. Our findings reveal​​ a tension between linguistic​​​‌ quality and benchmark performance:​ aggressive neural filtering improves​‌ fluency and coherence but​​ degrades benchmark scores, while​​​‌ deliberate late-stage contamination can​ recover competitive results at​‌ a moderate cost to​​ generation quality. These results​​​‌ shed light on how​ standard evaluation pipelines may​‌ inadvertently encourage benchmark leakage.​​ By additionally introducing controlled,​​​‌ harmless data poisoning, Gaperon​ provides a realistic testbed​‌ for future work on​​ robustness and safety, while​​​‌ establishing a solid empirical​ foundation for data-centric LLM​‌ research.

Finally, in the​​ context of Nathan Godey​​​‌ 's PhD, supervised by​ Benoît Sagot and Éric​‌ Villemonte De La Clergerie​​ , we investigated efficiency​​​‌ challenges at inference time,​ focusing on the memory​‌ bottleneck induced by the​​ Key-Value (KV) Cache in​​​‌ autoregressive models with long​ contexts 77, in​‌ collaboration with researchers from​​ the University of Edinburgh,​​​‌ the University of Rome​ and Miniml.AI. We introduce​‌ Q-Filters, a training-free KV​​ cache compression method based​​ on previously unexplored properties​​​‌ of query and key‌ vectors. By filtering less‌​‌ relevant key-value pairs using​​ a simple, context-agnostic projection,​​​‌ Q-Filters achieves strong compression‌ while remaining compatible with‌​‌ optimised attention implementations such​​ as FlashAttention. Experiments show​​​‌ that it matches or‌ outperforms existing methods across‌​‌ both retrieval and generation​​ tasks, enabling extreme compression​​​‌ levels with minimal performance‌ degradation.

8.1.2 Multimodal Language‌​‌ Modelling

In 2025 we​​ continued our activities in​​​‌ multimodal language modelling in‌ two different directions. A‌​‌ first research axis, in​​ collaboration with META through​​​‌ two CIFRE PhDs defended‌ in 2024 under Benoît‌​‌ Sagot 's academic supervision,​​ investigated how to tightly​​​‌ integrate language and speech‌ within a single generative‌​‌ framework. As a final​​ outcome of this collaboration,​​​‌ we introduced SPIRIT-LM, a‌ multimodal language model that‌​‌ treats text and speech​​ as a unified token​​​‌ stream 19. Starting‌ from a 7B-parameter pretrained‌​‌ text language model, we​​ continuously train on mixed​​​‌ sequences of text and‌ discretised speech units, interleaved‌​‌ at the word level.​​ This design enables the​​​‌ model to seamlessly switch‌ between modalities, combining the‌​‌ semantic reasoning strengths of​​ text language models with​​​‌ the expressive richness of‌ speech. Beyond a base‌​‌ version relying on phonetic​​ units, an expressive variant​​​‌ incorporates pitch and style‌ information, allowing the model‌​‌ to capture prosody and​​ expressivity. SPIRIT-LM demonstrates strong​​​‌ few-shot cross-modal generalisation, supporting‌ tasks such as ASR,‌​‌ TTS and speech classification​​ within a single architecture.​​​‌

A second, complementary research‌ axis addresses the data‌​‌ foundations required for scalable​​ and inclusive text-image multimodal​​​‌ learning, in the context‌ of Matthieu Futeral-Peter 's‌​‌ PhD thesis, defended in​​ 2025, and co-supervised by​​​‌ Rachel Bawden and Benoît‌ Sagot for ALMAnaCH and‌​‌ by Cordelia Schmid for​​ WILLOW (another research team​​​‌ from the Inria Paris‌ research centre, specialised in‌​‌ computer vision). While prior​​ work has shown how​​​‌ combining text and images‌ can be useful, including‌​‌ our own work on​​ the use of the​​​‌ image modality to improve‌ MT 107, existing‌​‌ multimodal (text-image) training corpora​​ are either private or​​​‌ limited to English, and‌ generally limited to captioning‌​‌ data. To overcome this​​ bottleneck, we developed and​​​‌ released mOSCAR, the first‌ large-scale multilingual and multimodal‌​‌ document corpus crawled from​​ the web 33.​​​‌ mOSCAR spans 163 languages‌ and combines text with‌​‌ over a billion images,​​ enabling research on multimodal​​​‌ models for a truly‌ global set of languages.‌​‌ Through extensive filtering and​​ evaluation, we ensure data​​​‌ quality and safety. To‌ prove mOSCAR's utility, we‌​‌ train a multilingual OpenFlamingo​​ from a Gemma-2B language​​​‌ model on a subset‌ of mOSCAR and captioning‌​‌ data derived from the​​ LAION-400M corpus and translated​​​‌ using NLLB. We compare‌ against a similar model‌​‌ trained on captioning data​​ only and show we​​​‌ obtain a strong boost‌ in few-shot learning, confirming‌​‌ previous findings for English.​​ Further results specifically for​​​‌ translation are mentioned in‌ the dedicated section (Section‌​‌ 8.3.5).

8.1.3 Code​​ Generation

In 2025 we​​​‌ continued our collaboration with‌ META on code generation‌​‌ with LLMs, in the​​​‌ context of Pierre Chambon​ 's CIFRE PhD, under​‌ Benoît Sagot 's academic​​ supervision. Our work this​​​‌ year has focused on​ advancing the evaluation and​‌ training of language models​​ for code generation, with​​​‌ particular emphasis on algorithmic​ reasoning and computational constraints.​‌ We developed and released​​ BigO(Bench) 72, a​​​‌ large-scale benchmark specifically designed​ to assess whether generative​‌ models can understand and​​ produce code that satisfies​​​‌ explicit time and space​ complexity requirements. Unlike prior​‌ evaluations that largely ignore​​ complexity, BigO(Bench) combines a​​​‌ principled profiling-based framework to​ infer algorithmic complexity of​‌ Python programs with a​​ rich dataset of more​​​‌ than 3,000 problems and​ over one million annotated​‌ solutions derived from competitive​​ programming. Experimental results reveal​​​‌ a nuanced picture: while​ modern token-space reasoning models​‌ excel at producing functionally​​ correct code, they often​​​‌ struggle to reliably meet​ complexity constraints, suggesting limited​‌ generalisation beyond training-time rewards.​​

Building on these findings,​​​‌ we have initiated work​ on reinforcement learning methods​‌ with the aim of​​ improving code generation quality​​​‌ and robustness under diverse​ objectives. In particular, we​‌ are exploring post-training with​​ Group Relative Policy Optimisation​​​‌ (GRPO) 123, an​ efficient RL approach that​‌ replaces value models with​​ Monte Carlo estimates. Our​​​‌ ongoing work investigates GRPO​ in a multi-task setting,​‌ where tasks are stochastically​​ sampled and training is​​​‌ organised through a global​ pool of workers and​‌ trainers. This line of​​ research lays the groundwork​​​‌ for more stable, scalable,​ and flexible RL-based optimisation​‌ of code-generating language models.​​

8.2.1 LLMs'​‌ Ability to Understand the​​ Concept of Persuasiveness

Alisa​​​‌ Barkar completed her PhD​ in December 2025. She​‌ was co-supervised by Chloé​​ Clavel , Mathieu Chollet​​​‌ (associate Professor at the​ University of Glasgow) and​‌ Beatrice Biancardi (Associate Professor​​ at CESI). During her​​​‌ last year, we investigate​ the application of LLMs​‌ for assessing public speaking​​ (PS) by predicting persuasiveness.​​​‌ We propose a novel​ framework where LLMs evaluate​‌ criteria derived from educational​​ literature and feedback from​​​‌ PS coaches, offering new​ interpretable textual features. We​‌ demonstrate that persuasiveness predictions​​ of a regression model​​​‌ with the new features​ achieve a Root Mean​‌ Squared Error (RMSE) of​​ 0.6, underperforming approach with​​​‌ hand-crafted lexical features (RMSE​ 0.51) and outperforming direct​‌ zero-shot LLM persuasiveness predictions​​ (RMSE of 0.8). Furthermore,​​​‌ we find that only​ LLM-evaluated criteria of language​‌ level is predictable from​​ lexical features (F1-score of​​​‌ 0.56), disapproving relations between​ these features. Based on​‌ our findings, we criticise​​ the abilities of LLMs​​​‌ to analyse PS accurately.​ To ensure reproducibility and​‌ adaptability to emerging models,​​ all source code and​​​‌ materials are publicly available​ on GitHub. This work​‌ was published at an​​ international conference (ICAART) 25​​​‌ and a national conference​ (TALN) 53.

8.2.2​‌ Evaluating Linguistic Diversity of​​ LLM outputs

Yanzhu Guo​​​‌ completed her PhD on​ automatic text evaluation in​‌ 2025 under the supervision​​ of Chloé Clavel .​​ Building on her earlier​​​‌ work on diversity decline‌ when recursively training language‌​‌ models on generated text​​ 108, we propose​​​‌ a comprehensive framework for‌ evaluating LLMs from multiple‌​‌ linguistic diversity perspectives 17​​. Using this framework,​​​‌ we benchmark six state-of-the-art‌ LLMs across lexical, syntactic,‌​‌ and semantic diversity on​​ five NLG tasks, with​​​‌ an in-depth case study‌ on syntactic diversity. We‌​‌ further analyse how model​​ scale, pretraining data, instruction​​​‌ tuning, decoding strategies, and‌ quantisation affect diversity. Our‌​‌ results show that, despite​​ strong task-solving abilities, current​​​‌ LLMs fall short of‌ human linguistic richness, particularly‌​‌ for creative tasks such​​ as story generation. Notably,​​​‌ instruction tuning increases lexical‌ diversity but constrains syntactic‌​‌ and semantic diversity, suggesting​​ reduced expressive flexibility. These​​​‌ findings highlight a risk‌ of homogenisation in LLM-generated‌​‌ content and highlights the​​ need to prioritise linguistic​​​‌ diversity alongside traditional performance‌ metrics.

8.2.3 Political Analysis‌​‌

In this research direction,​​ led by Carlo Santagiustina​​​‌ , we provide large-scale‌ comparative evidence on how‌​‌ citizens across the European​​ Union mobilise for e-petitions​​​‌ through Social Media, and‌ how such mobilisation is‌​‌ shaped by ideological orientations​​ toward political elites and​​​‌ economic policy 87.‌ By analysing over 1.8‌​‌ million X/Twitter profiles across​​ five EU member states​​​‌ (Germany, France, Italy, the‌ Netherlands, and Poland), we‌​‌ model the likelihood that​​ users share petition-related content​​​‌ as a function of‌ their individual ideological positions,‌​‌ the average ideological positions​​ of the parties they​​​‌ follow, the distance between‌ the two, and a‌​‌ set of pseudo-demographic and​​ user-activity controls. Our findings​​​‌ show that online petition‌ mobilisation in Europe is‌​‌ primarily driven by individual-level​​ ideological attitudes, rather than​​​‌ by the ideological orientation‌ of parties or political‌​‌ groups as such. A​​ key result is that​​​‌ mobilisation for e-petitions often‌ reflects perceived gaps in‌​‌ representation rather than coordinated​​ party-led engagement. While the​​​‌ average ideological positions of‌ parties followed by users‌​‌ do not significantly increase​​ the likelihood of petition​​​‌ sharing, ideological distance from‌ those party positions (especially‌​‌ along the anti-elite dimension)​​ substantially raises the probability​​​‌ of mobilisation. This indicates‌ e-petitions tend to function‌​‌ as channels for citizens​​ to express political discontent​​​‌ when they feel insufficiently‌ represented by actors they‌​‌ otherwise follow or identify​​ with, highlighting their role​​​‌ as indicators of representational‌ strain within European democracies.‌​‌ Our work highlights the​​ strategic value of analysing​​​‌ Social Media-based mobilisation as‌ a complementary tool for‌​‌ democratic governance at national​​ and EU levels. Patterns​​​‌ of petition sharing, especially‌ when concentrated in particular‌​‌ ideological groups, can serve​​ as early signals of​​​‌ rising political discontent not‌ yet visible through electoral‌​‌ outcomes or traditional surveys.​​ Monitoring e-petition mobilisation on​​​‌ Social Media can help‌ anticipate contention, improve responsiveness‌​‌ to citizens' needs and​​ concerns, and better align​​​‌ institutional engagement with how‌ Europeans express political preferences‌​‌ and dissatisfaction online.

8.2.4​​ Cultural Adaptation of LLMs​​​‌

In the context of‌ Gabrielle Le Bellier 's‌​‌ PhD, supervised by Chloé​​ Clavel and Benoît Sagot​​​‌ , we explored the‌ lack of cultural commonsense‌​‌ knowledge of LLMs. LLMs​​​‌ struggle to answer cultural​ questions about daily life,​‌ such as “What is​​ a common snack for​​​‌ preschool kids in [country]?”,​ where we focus on​‌ a culture at a​​ country scale. They show​​​‌ disparate performances between high-resource​ cultures and low-resource ones,​‌ for instance with high​​ accuracies for the United​​​‌ Kingdom and low accuracies​ for North Korea or​‌ Ethiopia. By using soft​​ prompt methods on OLMo-7B-Instruct,​​​‌ we showed that the​ model's ability to answer​‌ cultural commonsense questions can​​ be improved. These methods​​​‌ require few computational resources​ by tuning only a​‌ few parameters added to​​ the frozen model (approximately​​​‌ 0.001% for prompt-tuning and​ 1% for prefix-tuning). We​‌ fine-tuned the soft prompts​​ on cultural multi-choice question​​​‌ answering and obtained one​ soft prompt per country.​‌ We observed cultural proximity​​ by evaluating a prompt​​​‌ on other cultures' questions.​ For instance, the Spain​‌ prompt-tuned model is better​​ at answering questions about​​​‌ Mexico than questions about​ North Korea. Moreover, we​‌ argue that we can​​ interpret similarities between soft​​​‌ prompts to deduce cultural​ proximity. This work was​‌ co-funded by Benoît Sagot​​ 's PRAIRIE chair position​​​‌ and Chloé Clavel 's​ PRME SINNet ANR project,​‌ in collaboration with the​​ “Action exploratoire” SaLM.

8.2.5​​​‌ Argumentation in Political Debates​

In the context of​‌ Cecilia Graiff 's PhD,​​ funded by Benoît Sagot​​​‌ 's PRAIRIE chair position​ and co-supervised by Benoît​‌ Sagot and Chloé Clavel​​ in collaboration with the​​​‌ “Action Exploratoire” SaLM (Socially​ Aware Language Models), we​‌ benchmarked the performance of​​ language models on cross-lingual​​​‌ and cross-country argument component​ detection, focusing on political​‌ data from the US​​ and France. We considered​​​‌ natural language arguments as​ composed by claims and​‌ premises, following the most​​ reliable approaches in argument​​​‌ mining. To overcome the​ lack of data in​‌ this field, we introduced​​ FrenchPolArg, a corpus of​​​‌ argumentative political discourse in​ French, and we automatically​‌ translated and projected the​​ annotations of ElecDeb60to20, a​​​‌ corpus of US presidential​ debates annotated for argument​‌ components and relations. We​​ benchmarked three different cross-lingual​​​‌ and cross-country pipelines, and​ we compared their results​‌ to find the best-performing​​ one, namely model transfer.​​​‌ We obtained promising results​ to be integrated in​‌ semi-automatic annotation workflows to​​ reduce the time and​​​‌ cost of annotations. Moreover,​ we are launching an​‌ annotation campaign with expert​​ annotators from SciencesPo Paris,​​​‌ in order to obtain​ a ground truth and​‌ continue our experiments. In​​ the future, this PhD​​​‌ aims to explore the​ abilities of language models​‌ to understand reasoning patterns.​​

8.3.1 MT​‌ Benchmarking and Shared Tasks​​

As in previous years,​​​‌ Rachel Bawden was a​ member of the organising​‌ committee of the general​​ MT shared task at​​​‌ the main conference in​ MT (WMT) 82,​‌ 40 (she has previously​​ also been part of​​ the committee for the​​​‌ biomedical task, but which‌ did not take place‌​‌ this year). The purpose​​ of this shared task​​​‌ is the evaluation of‌ models submitted by participants‌​‌ on the automatic translation​​ of texts from multiple​​​‌ domains, namely news, literary,‌ speech and social. The‌​‌ shared task has evolved​​ over time, and this​​​‌ year the focus was‌ particularly on document-level MT.‌​‌ Rachel Bawden assisted notably​​ with the social media​​​‌ data, which this year‌ included screenshots of threads‌​‌ from Mastodon in order​​ to enable the participation​​​‌ of mulitmodal text-image models.‌ Rachel Bawden and Benoit‌​‌ Sagot also participated in​​ the test suite shared​​​‌ task, with an updated‌ version of the RoCS-MT‌​‌ (Robust Challenge Set for​​ MT) challenge set 27​​​‌, a parallel set‌ designed for the evaluation‌​‌ of MT of highly​​ non-standard user-generated content (UGC),​​​‌ which is detailed more‌ in Section 8.3.3 below.‌​‌ Rasul Jasir Dent was​​ also member of the​​​‌ organising committee for the‌ first shared taks in‌​‌ Creole MT, at the​​ same conference, which encouraged​​​‌ submissions of MT systems‌ and datasets for MT‌​‌ involving Creole languages.

In​​ terms of benchmarking, Rachel​​​‌ Bawden continued her participation‌ in an initiative led‌​‌ by Jesujoba Alabi (a​​ previous engineer in the​​​‌ team and now a‌ PhD student at Saarland‌​‌ University) on the creation​​ of a document-level domain-specific​​​‌ (health and tech) corpus‌ for African languages, AfriDoc-MT.‌​‌ The benchmark was used​​ to evaluate LLM performance​​​‌ at the document level,‌ showing that document-level MT‌​‌ remains challenging for these​​ languages. The work was​​​‌ finalised and published at‌ EMNLP 2025 22.‌​‌

8.3.2 MT for Open​​ Science

We continued our​​​‌ work on MT for‌ scientific documents in the‌​‌ context of the MaTOS​​ (Machine Translation for Open​​​‌ Science) ANR project 26‌, led by François‌​‌ Yvon (CNRS). There is​​ a specific focus on​​​‌ English–French and French–English translation,‌ since the high quality‌​‌ of MT for these​​ directions allows us to​​​‌ work on more high-level‌ challenges such as term‌​‌ translations, consistency and document​​ coherence. Two PhD students​​​‌ are currently co-supervised by‌ Rachel Bawden jointly with‌​‌ François Yvon in the​​ projet: Ziqian Peng (recruited​​​‌ by the CNRS) on‌ document-level MT for scientific‌​‌ documents and Nicolas Dahan​​ on the evaluation of​​​‌ MT.

Following her work‌ on improving translation for‌​‌ longer documents last year​​ (with full document context),​​​‌ this year Ziqian Peng‌ has mainly been focusing‌​‌ on extracting relevant local​​ and distant contexts to​​​‌ help translation. The main‌ motivation for the work‌​‌ is that context can​​ be useful, but is​​​‌ typically sparse (not all‌ context will be useful)‌​‌ and it can be​​ inefficient and sometimes detrimental​​​‌ to include full document‌ context when translating. It‌​‌ may be beneficial to​​ be able to extract​​​‌ context from the left‌ context in a document‌​‌ (i.e. the previously translated​​ sentences) that can either​​​‌ be local or more‌ distant, without having to‌​‌ include full document context.​​ With the proposed self-RAMT​​​‌ approach (for retrieval-augmented MT‌ framework), it is possible‌​‌ to dynamically extract contexts​​​‌ from the source text​ and the associated translation.​‌ We carried out experiments​​ on three LLMs, considering​​​‌ three different criteria to​ select relevant contexts to​‌ include: (i) cosine similarity​​ using LaBSE 105,​​​‌ a variant of BM25L​ 110 and PMI (to​‌ identify contexts that are​​ most likely to make​​​‌ a difference). We carry​ out experiments on TED​‌ talks as well as​​ parallel scientific articles, considering​​​‌ three translation directions. Our​ results show that integrating​‌ distant contexts does improve​​ translation quality as measured​​​‌ by reference-based scores and​ consistency metrics.

In addition​‌ to models fine-tuned for​​ the project (on scientific​​​‌ documents from Natural Language​ Processing (NLP) and Earth​‌ and Planetary Sciences), described​​ in 86, we​​​‌ have also been contributing​ to the collection and​‌ processing of additional resources​​ for training and evaluation​​​‌ of MT models for​ scientific documents. In a​‌ paper currently under submission,​​ we describe the collection​​​‌ of English–French parallel corpora​ for NLP and Earth​‌ and Planetary Sciences, including​​ train, development and test​​​‌ sets and covering 14k​ abstracts and 104 full-length​‌ articles. The parallel data​​ comes from a variety​​​‌ of sources and includes​ huamn-produced translations as well​‌ as automatic translations to​​ fill the gaps where​​​‌ no reference translations exist.​ We test the performance​‌ of eight MT systems​​ on our test sets​​​‌ and find that fine-tuning​ on our data reduces​‌ the gap between open​​ LLM-based systems and commercial​​​‌ ones. We also find​ that the performance of​‌ recent LLMs can worsen​​ when translating full articles​​​‌ instead of translating them​ at the paragraph level.​‌

In the context of​​ Nicolas Dahan 's PhD,​​​‌ we continued to look​ at ways of evaluating​‌ document-level consistency and coherence​​ in translations. We finalised​​​‌ the MetaDocEval dataset, comprising​ contrastive test examples of​‌ correct and incorrect translations.​​ The incorrect samples were​​​‌ automatically generated to contain​ specific errors designed to​‌ stress test metrics on​​ potential errors of document​​​‌ MT systems (e.g. truncating​ translations, shuffling sentences, changing​‌ pronouns) and to see​​ the effect on metrics​​​‌ of phenomena to which​ we would expect them​‌ to be robust (splitting​​ sentences, use of synonyms).​​​‌ Testing on English-to-French, Spanish​ and German, we evaluated​‌ several MT metrics, using​​ multiple segmentation strategies and​​​‌ span lengths. Our experiments​ showed that some metrics​‌ highly sensitive to local​​ changes under-estimate long-range disruptions,​​​‌ and vice versa and​ that certain perturbations (e.g.​‌ related to coherence and​​ repetition) reveal shortcomings in​​​‌ metrics that would appear​ robust under sentence-level evaluations.​‌ The paper is also​​ under submission.

We have​​​‌ also turned our attention​ to the evaluation of​‌ term use in translation,​​ since this is a​​​‌ highly important aspect of​ scientific document translation; terms​‌ must not only be​​ translated correctly but they​​​‌ must also be translated​ in a coherent way​‌ in the document (this​​ can correspond to consistency​​​‌ in some cases, but​ it also corresponds to​‌ the appropriate use of​​ terminological variants such as​​​‌ acronyms, synonyms, and reductions).​ With the aim of​‌ modelling terminological use within​​ documents, Éric de la​​ Clergerie has been leading​​​‌ research into tools to‌ identify terms and their‌​‌ variants, with significant developments​​ made to the Concordancer​​​‌ tool 7.1.16, described‌ in another submitted paper.‌​‌ The tool uses a​​ predefined glossary to detect​​​‌ terms within documents and‌ then detects variants based‌​‌ on variation of the​​ syntactic structure of terms,​​​‌ their morphological variants, synonyms‌ and reductions. It also‌​‌ detects chains of variants​​ within documents, including those​​​‌ from the glossary and‌ those detected dynamically. The‌​‌ tool can then in​​ turn be used to​​​‌ enrich terminologies once it‌ has been applied to‌​‌ scientific documents, which we​​ do for the NLP​​​‌ and Earth and Planetary‌ Sciences fields. We also‌​‌ carried out preliminary experiments​​ in the medical domain.​​​‌

The Concordancer is being‌ used in ongoing research‌​‌ to model terminological use​​ in human and automatic​​​‌ translations as part of‌ Nicolas Dahan 's PhD,‌​‌ detecting terms and their​​ variants to identify what​​​‌ the differences in use‌ are and what constitutes‌​‌ a correct use of​​ a term, especially in​​​‌ the context of the‌ document. The aim is‌​‌ to use this analysis​​ to develop an automatic​​​‌ metric that can be‌ used to evaluate term‌​‌ use in document-level MT.​​

One of the final​​​‌ outcomes of the project‌ will hopefully be an‌​‌ online tool to allow​​ authors of scientific articles​​​‌ to automatically translate their‌ abstracts and then post-edit‌​‌ them when submitting papers​​ to the HAL open​​​‌ archive (at least for‌ Inria's instance of HAL).‌​‌ Panagiotis Tsolakis is currently​​ working on an implementation​​​‌ to do just that,‌ involving translating abstracts with‌​‌ our dedicated trained models​​ (and testing different scenarios​​​‌ including zero-shot, few-shot translating‌ with example selection) and‌​‌ then applying multiple verification​​ steps to ensure a​​​‌ reasonable quality of automatic‌ translation.

8.3.3 MT for‌​‌ Non-Standard Data

MT performance​​ for high-resource, edited texts​​​‌ has improved dramatically over‌ recent years, and LLMs‌​‌ have also reached state-of-the-art​​ levels. However, challenges still​​​‌ remain for texts that‌ display variation with respect‌​‌ to standard language varieties.​​ The case of translation​​​‌ for low-resource dialects is‌ discussed in the next‌​‌ section (Section 8.3.4).​​ In this section, we​​​‌ cover this year's contributions‌ in the translation of‌​‌ contemporary user-generated content (UGC)​​ and in the resolution​​​‌ of abbreviations in Mediaeval‌ manuscripts, framed as a‌​‌ translation task.

8.3.4 Low-resource​​​‌ and Dialectal MT

Another‌ major challenge of MT‌​‌ is developing approaches to​​ handle scenarios where there​​​‌ is typically little good‌ quality data (monolingual or‌​‌ parallel) available, and where​​ current LMs and MT​​​‌ models therefore underperform. We‌ have developed this research‌​‌ direction in 2025, with​​ several different projects tackling​​​‌ a range of scenarios,‌ including translation from and‌​‌ into under-represented languages but​​ which do appear in​​​‌ training data, translation from‌ and to very low-resource‌​‌ languages (requiring explicit information​​ from grammar books to​​​‌ carry out translation) and‌ translation involving low-resource dialect‌​‌ continuums.

8.3.5 Multimodal​​​‌ MT

A second PhD‌ in MT was defended‌​‌ in 2025: Matthieu Futeral-Peter​​ successfully defended his PhD​​​‌ on multimodal language modelling‌ (with a focus on‌​‌ MT), supervised by Rachel​​ Bawden , Benoît Sagot​​​‌ and Cordelia Schmid from‌ the WILLOW project-team.28‌​‌

Two of his previously​​ started projects were finalised​​​‌ and published in 2025‌ at international conferences: mOSCAR‌​‌ in ACL Findings 33​​ and ZeroMMT in NAACL​​​‌ Findings 32.

mOSCAR,‌ also mentioned in Section‌​‌ 8.1.2 for aspects on​​ multimodal language modelling, is​​​‌ a multimodal version of‌ the OSCAR corpus 117‌​‌, comprising web documents​​ and associated images, which​​​‌ can be used to‌ train multimodal language models.‌​‌ The corpus has the​​ strength of being highly​​​‌ multilingual, covering 163 languages‌ and therefore being a‌​‌ potentially useful source for​​ inherently multilingual tasks such​​​‌ as MT. As described‌ previously, we trained a‌​‌ multimodal language model on​​ the data (in addition​​​‌ to more traditional caption-like‌ resources), and one of‌​‌ the downstream tasks tested​​ was MT. We showed​​​‌ that training on the‌ additional interleaved data from‌​‌ mOSCAR29 results in​​ improved MT performance on​​​‌ the traditional Multi30k test‌ set 104 (the model‌​‌ only trained on captioning​​ data cannot translate at​​​‌ all) and is also‌ able to effectively use‌​‌ images to disambiguate ambiguous​​ texts in our CoMMuTE​​​‌ contrastive dataset 107.‌

ZeroMMT 32 is a‌​‌ method to train multimodal​​ MT systems without requiring​​​‌ fully supervised multimodal parallel‌ data (i.e., sentence pairs‌​‌ with corresponding images). The​​​‌ approach involves adapting a​ strong text-only MT model​‌ using a joint training​​ objective: visually conditioned masked​​​‌ language modelling and a​ Kullback-Leibler divergence loss that​‌ encourages the MMT model's​​ outputs to remain consistent​​​‌ with the original MT​ model while integrating visual​‌ cues. This approach allows​​ for effective learning from​​​‌ multimodal English data alone,​ making it applicable to​‌ language pairs with no​​ existing multimodal training corpora.​​​‌ ZeroMMT demonstrated disambiguation performance​ close to state-of-the-art multimodal​‌ MT models trained on​​ fully supervised data, as​​​‌ evaluated on standard MMT​ benchmarks and our contrastive​‌ evaluation set CoMMuTE. To​​ further validate its generalisation​​​‌ capabilities, we extended CoMMuTE​ to three new languages:​‌ Arabic, Russian, and Chinese.​​ Additionally, we introduced an​​​‌ inference-time trade-off mechanism using​ classifier-free guidance, allowing for​‌ a controlled balance between​​ translation fidelity and disambiguation​​​‌ strength without requiring additional​ data.

8.4.1 Identifying Common Examples​ in Spanish Dialects in​‌ Hate-Speech Contexts

Our previous​​ research on multilingual zero-shot​​​‌ hate speech detection 113​ highlighted the cultural gap​‌ in cross-lingual settings. For​​ instance, the Spanish word​​​‌ puta `prostitute', is often​ used as a non-offensive​‌ intensifier but was flagged​​ as offensive by models​​​‌ trained on an English​ dataset that focuses on​‌ misogyny. While such cultural​​ differences across languages are​​​‌ increasingly studied, their impact​ on dialects of the​‌ same language remains unclear.​​ After our work on​​​‌ hate speech contrasting different​ varieties of Spanish spoken​‌ either in South America​​ or in Spain 99​​​‌, we extended our​ approach, this time focusing​‌ on the Cuban Spanish​​ variant and the identification​​​‌ of cross-dialect common examples.​

As mentioned above, variations​‌ in languages across geographic​​ regions or cultures are​​​‌ crucial to addressing biases​ in NLP systems designed​‌ for culturally sensitive tasks,​​ such as hate speech​​​‌ detection or dialogue with​ conversational agents. In languages​‌ such as Spanish, where​​ varieties can significantly overlap,​​​‌ many examples can be​ valid across varieties, which​‌ we refer to as​​ common examples. Ignoring these​​​‌ examples may cause misclassifications,​ reducing model accuracy and​‌ fairness. Accounting for these​​ common examples is therefore​​​‌ essential to improving the​ robustness and representativeness of​‌ NLP systems trained on​​ such data. We used​​ training dynamics to automatically​​​‌ detect common examples or‌ errors in existing Spanish‌​‌ datasets. We demonstrated the​​ effectiveness of using predicted​​​‌ label confidence for our‌ Datamaps 125 implementation for‌​‌ the identification of hard-to-classify​​ examples, especially common examples,​​​‌ enhancing model performance in‌ a variety of identification‌​‌ tasks. Besides the original​​ use of the Datamaps​​​‌ for this task, an‌ interesting outcome lies in‌​‌ the introduction of a​​ Cuban Spanish Variety Identification​​​‌ dataset with common example‌ annotations developed to facilitate‌​‌ more accurate detection of​​ Cuban and Caribbean Spanish​​​‌ varieties. To our knowledge,‌ this is the first‌​‌ dataset focused on identifying​​ the Cuban, or any​​​‌ other Caribbean, Spanish variety.‌ This work was published‌​‌ in the VARDIAL specialised​​ workshop, colocated with COLING​​​‌ 2025 23.

8.4.2‌ Quantifying Geographically-Informed Sociocultural Biases‌​‌

LLMs exhibit inequalities with​​ respect to various cultural​​​‌ contexts. Most prominent open-weights‌ models are trained on‌​‌ Global North data and​​ show prejudicial behaviour towards​​​‌ other cultures. Moreover, there‌ is a notable lack‌​‌ of resources to detect​​ biases in non-English languages,​​​‌ especially from Latin America‌ (Latam), a continent containing‌​‌ various cultures, even though​​ they share a common​​​‌ cultural ground. This aspect‌ is especially import ant‌​‌ in the context of​​ the AeX SaLM and​​​‌ the ongoing collaboration with‌ Inria Chile. As a‌​‌ starting point, we proposed​​ to evaluate how culturally​​​‌ biased were currrent LLMs‌ towards Latam. To do‌​‌ so we leveraged the​​ content of Wikipedia, the​​​‌ structure of the Wikidata‌ knowledge graph, and expert‌​‌ knowledge from social science​​ in order to create​​​‌ a dataset of Question-Answer‌ pairs, based on the‌​‌ different popular and social​​ cultures of various Latin​​​‌ American countries. We created‌ a database of around‌​‌ 20k questions and associated​​ answers extracted from 20k​​​‌ Wikipedia articles, transformed into‌ a multiple-choice questions in‌​‌ Spanish and Portuguese, in​​ turn translated to English.​​​‌ We use this dataset‌ to quantify the degree‌​‌ of knowledge of various​​ LLMs and find (i)​​​‌ a discrepancy in performances‌ between the Latam countries,‌​‌ some being easier than​​ others for the majority​​​‌ of the models, (ii)‌ that the models perform‌​‌ better in their original​​ language, and (iii) that​​​‌ Iberian Spanish culture is‌ better known than the‌​‌ Latam one.

This work​​ was carried out by​​​‌ Yannis Karmim and Djamé‌ Seddah in collaboration with‌​‌ Luis Martí (Inria Chile)​​ and Valentin Barriere (University​​​‌ of Santiago) under the‌ Inria Chile-Inria collaboration framework‌​‌ and the AeX SaLM.​​ A paper will be​​​‌ published as part of‌ the proceedings of the‌​‌ Multilingual Multicultural Evaluation workshop​​ collocated with EACL 2026.​​​‌

8.4.3 The CounteR Dataset‌

The first half of‌​‌ the CounteR project faced​​ an almost fatal challenge,​​​‌ namely the lack of‌ availability of a multilingual‌​‌ radical content dataset covering​​ the project's target ideology​​​‌ spectrum (from white-supremacist to‌ Jihadism). This is why‌​‌ we had to first​​ develop our multitask learning​​​‌ architecture on adjacent domains‌ such as hate speech‌​‌ before being able to​​ focus on real radical​​​‌ content suitable for the‌ law enforcement agencies that‌​‌ were part of the​​​‌ project and designed to​ be our end-users. Such​‌ data was made available​​ to us by an​​​‌ EU-external third-party contractor, validated​ by the European Commission​‌ project officer, over an​​ 18-month period during the​​​‌ second half of the​ project. The final dataset​‌ covers 12 languages, some​​ of which were actually​​​‌ provided by the LEAs​ themselves. ALMAnaCH was responsible​‌ for the whole NLP​​ work package (“Data Analytics​​​‌ for Detecting Radical Content”),​ which spanned the entire​‌ duration of the project.​​ Due to the UE​​​‌ classified nature of our​ core deliverables (D4.3 “NLP​‌ Basic Features” and D4.4​​ “transfer learning for NLP”)​​​‌ we could not present​ our results on cross-lingual​‌ transfer and out-of-domain scenarios​​ on radical content detection​​​‌ here. As per the​ consortium agreement, the only​‌ data that could be​​ released would have to​​​‌ be fully GDPR-compliant, entailing​ a perfect pseudo-anonymisation process​‌ as demonstrated in 109​​. Therefore, given the​​​‌ extent of the work​ needed for this task,​‌ we focused on a​​ subset of languages comprising​​​‌ French, Arabic and English.​ The idea was to​‌ not only pseudo-anonymise the​​ data-set in these three​​​‌ languages but also, while​ doing so, to keep​‌ all socio-demographic information and​​ cultural traits that could​​​‌ have been linked to​ named-entities tied to the​‌ highly specific radicalisation domain.​​ The goal was two-fold:​​​‌ (i) for a classifier​ trained on the new​‌ data to match the​​ performance of a classifier​​​‌ trained on the original​ data set and (ii)​‌ to allow the redistribution​​ of this dataset while​​​‌ being fully compliant to​ the GPDR.

It is​‌ important to note that​​ strong anonymisation is of​​​‌ course necessary, but overly​ aggressive approaches can remove​‌ important social and contextual​​ signals, such as usernames,​​​‌ locations, or references to​ public events, which are​‌ often crucial for detecting​​ radical content 118.​​​‌ As a result, balancing​ privacy protection and data​‌ usefulness is a major​​ challenge. This is why​​​‌ we developed a manual​ pseudonymisation method tailored to​‌ our radicalisation dataset. Our​​ approach preserves key semantic​​​‌ information while protecting user​ identities, allowing models trained​‌ on the processed data​​ to achieve performance comparable​​​‌ to that of the​ original data. The process​‌ includes careful annotation of​​ named entities and selective​​​‌ preservation of well-known public​ figures and events. This​‌ resulted in a multilingual,​​ semantically rich, and privacy-compliant​​​‌ dataset in English, French,​ and Arabic, which we​‌ made freely available for​​ research 120.

8.4.4 Inferring​​ Sociological Variation in Algorithmic​​​‌ Text Classification

Following our​ work on the CounteR​‌ Dataset, Sofia Imbert de​​ Tremiolles has recently started​​​‌ her PhD, funded by​ the AeX SaLM and​‌ co-supervised by Djamé Seddah​​ and Jean-Philippe Cointet (Sciences​​​‌ Po). A first axis​ of work will be​‌ to leverage existing longitudinal​​ data from Twitter and​​​‌ detailed information about the​ socioeconomic status and political​‌ preferences of nearly 7,000​​ individuals that are available​​​‌ through an 800 million​ French tweets dataset (SoSweet​‌ corpus).

If robust discrepancies​​ are observed in model​​​‌ quality between socioeconomic, geographical,​ or political backgrounds, it​‌ is possible that these​​ variations reflect an underlying​​​‌ social phenomenon, or that​ it is simply an​‌ undesirable mistreatment of certain​​ texts by the algorithm​​​‌ itself. In that case,​ we will develop methodology​‌ to systematically test which​​ of these possibilities applies​​​‌ in each case.

In​ all cases, the SoSweet​‌ dataset will be used​​ to test the effectiveness​​​‌ of various NLP algorithms,​ conditional on the authors'​‌ social, geographic, and political​​ context. Key questions are:​​​‌

• How does the reliability​ of toxicity detection vary​‌ across social classes?
• Is​​ stance detection more sensitive​​​‌ for right-wing individuals than​ left-wing individuals?
• Does a​‌ cultural gap between geographic​​ areas or different social​​​‌ groups make it harder​ to detect sexist content?​‌

8.4.5 Abusive Language Detection​​ in Online Conversations

In​​​‌ the context of Célia​ Nouri 's PhD thesis,​‌ under the supervision of​​ Chloé Clavel and Jean-Philippe​​​‌ Cointet, we contributed to​ the detection of abusive​‌ language in online conversations​​ by leveraging contextual cues​​​‌ from the conversation history​ (i.e. previous comments in​‌ the online thread) 45​​. Our research demonstrated​​​‌ that abusive language detection​ models can be significantly​‌ improved by integrating conversational​​ features that capture both​​​‌ the content and topology​ of discussions. In particular,​‌ we introduced a graph-based​​ framework that models social​​​‌ media conversations as structured​ graphs, enabling the propagation​‌ of contextual cues across​​ reply chains through the​​​‌ use of Graph Neural​ Networks. We compare our​‌ graph-based model with context-agnostic​​ and flattened context-aware baselines​​​‌ and demonstrate that our​ method provides substantial improvements​‌ in abusive language detection,​​ especially in cases where​​​‌ abusiveness is implicit and​ can only be inferred​‌ through extended interaction history.​​ This research was funded​​​‌ by Chloé Clavel 's​ SINNet ANR project, in​‌ collaboration with the Action​​ Exploratoire SaLM. The article​​​‌ was published at the​ Association for Computational Linguistics​‌ (ACL) and received the​​ Senior Area Chair Highlights​​​‌ award, a distinction granted​ to approximately 4% of​‌ the accepted papers, underscoring​​ both its methodological contribution​​​‌ and its impact on​ the field.

8.6.1 Repair Modelling​​

Regarding the scope of​​​‌ Anh Ha Ngo 's‌ PhD, funded by the‌​‌ Paris Île-de-France Région through​​ the DIM AI4IDF and​​​‌ under the supervision of‌ Chloé Clavel , Catherine‌​‌ Pelachaud (Sorbonne Université, ISIR)​​ and Nicolas Rollet (Telecom-Paris,​​​‌ on secondment in ALMAnaCH),‌ our research focuses on‌​‌ modelling conversational repair mechanisms​​ to improve human-agent interaction.​​​‌ Building upon our initial‌ linguistic analysis in the‌​‌ first year,114 we​​ published a novel multimodal​​​‌ approach for detecting human‌ repair initiation in the‌​‌ EMNLP 2025 paper, “Mm,​​ Wat?” Detecting Other-initiated Repair​​​‌ Requests in Dialogue 44‌. This work proposed‌​‌ a computational model that​​ integrates retrained text and​​​‌ audio embeddings with handcrafted‌ linguistic and prosodic features‌​‌ grounded in Conversation Analysis​​ literature. Specifically, the results​​​‌ suggested that prosodic cues,‌ such as pausing behaviours‌​‌ and voice intensity, synergise​​​‌ with linguistic patterns, such​ as coreference usage and​‌ repetition, play an essential​​ role in identifying when​​​‌ a user signals trouble​ and initiates a need​‌ for repair.

To expand​​ this research by evaluating​​​‌ the generalisability of repair​ detection across different languages​‌ and models, in the​​ second half of 2025,​​​‌ we submitted a study​ to LREC 2026 involving​‌ the annotation of other-initiated​​ repair in the French​​​‌ NOXI corpus. We assessed​ the capabilities of LLMs​‌ to serve as both​​ automated annotators and explainers​​​‌ for the disagreements in​ annotation between LLMs, human​‌ annotator and human experts.​​

8.6.2 Word Meaning Negotiation​​​‌

As part of Aina​ Garí Soler 's postdoctoral​‌ research, and towards the​​ end of the year​​​‌ in her capacity as​ an associated member, we​‌ continued our investigations on​​ word usage dynamics in​​​‌ dialogue. This work was​ supervised by Chloé Clavel​‌ and Matthieu Labeau (Telecom-Paris)​​ and funded by Chloé​​​‌ Clavel 's ANR PRME​ SINNet project. Our submission​‌ to the LRE journal​​ (Language Resources and Evaluation)​​​‌ on the NeWMe corpus​ is now available as​‌ a preprint 89 and​​ the corpus is available​​​‌ online. The NeWMe (Negotiating​ Word Meaning) corpus is​‌ a collection of annotated​​ Word Meaning Negotiation (WMN)​​​‌ sequences. WMNs are parts​ of conversations where a​‌ speaker raises a clarification​​ request or an objection​​​‌ about the usage of​ a specific word, prompting​‌ a metalinguistic discussion aimed​​ at clarifying or refining​​​‌ the word's meaning. The​ corpus contains various kinds​‌ of interactions (written, oral,​​ dyadic, and multi-party) and​​​‌ is annotated with WMN​ sequences, their components, and​‌ related phenomena.

NeWMe is​​ the first corpus of​​​‌ its kind, but it​ is of limited size,​‌ and we have worked​​ on developing methods for​​​‌ its semi-automatic extension. WMNs​ have three key components:​‌ the trigger, which is​​ a problematic word usage;​​​‌ the indicator, an utterance​ signaling a problem with​‌ a word's meaning; and​​ the negotiation—one or more​​​‌ conversational turns where participants​ address the misunderstanding or​‌ disagreement. We have specifically​​ explored the automatic detection​​​‌ of WMN indicators in​ conversation. Our proposed models​‌ have a much better​​ precision than previous regex-based​​​‌ identification methods, and show​ promise for accelerating the​‌ WMN identification process, but​​ there is still room​​​‌ for improvement and we​ identify relevant directions for​‌ future work. This work​​ resulted in an article​​​‌ accepted at Findings of​ EMNLP 2025 35.​‌

To better understand the​​ mechanisms by which a​​​‌ word can be misunderstood​ or disputed, we paid​‌ particular attention to the​​ trigger component. We conducted​​​‌ an extensive literature review,​ identifying linguistic factors that​‌ contribute to potentially problematic​​ word usages, as well​​​‌ as computational methods and​ data that facilitate their​‌ detection. This review connects​​ works from various disciplines​​​‌ and highlights areas for​ future research. The outcome​‌ of this work is​​ a survey that has​​​‌ been accepted to *SEM​ 2025 34.

Finally,​‌ we focused on scare​​ quoted word usages as​​​‌ examples of potentially problematic​ word usages that could​‌ trigger a WMN. Quotation​​ marks have a wide​​ range of uses in​​​‌ written language. One of‌ them is scare quoting:‌​‌ scare quotes mark irony,​​ distance, or disagreement about​​​‌ word meaning or lexical‌ choices. As such, they‌​‌ are explicit markers of​​ the writer's acknowledgment that​​​‌ a word may be‌ problematic and their anticipation‌​‌ of a possible misalignment​​ in meaning. We created​​​‌ the first annotated corpus‌ of quoted word and‌​‌ phrase usages focused on​​ the scare vs non-scare​​​‌ quotes distinction, and in‌ their use in conversation.‌​‌ This work is currently​​ under review at LREC​​​‌ 2026 and it will‌ contribute to the detection‌​‌ and characterisation of problematic​​ word usages.

8.7.1 Biomedical​​​‌ and Clinical Domains

During‌ her PhD, Lucie Chenain‌​‌ supervised by Chloé Clavel​​ and Anne-Catherine Bachoud-Levi (APHP,​​​‌ ENS, Inserm), introduced emoHD,‌ the first speech corpus‌​‌ designed to study emotional​​ expression and psychiatric symptoms​​​‌ in Huntington's Disease (HD)‌ 15. Addressing a‌​‌ major clinical challenge in​​ HD, this work investigated​​​‌ how emotions expressed in‌ speech relate to the‌​‌ severity of psychiatric symptoms,​​ including depression, irritability/aggressivity, apathy,​​​‌ and obsessive/compulsive symptoms. HD‌ participants first underwent a‌​‌ clinical assessment in which​​ neurologists evaluated these psychiatric​​​‌ symptoms using standardised measures.‌ They then completed a‌​‌ recorded speech interview conducted​​ by a neuropsychologist. These​​​‌ recordings were then emotionnally‌ annotated using a three-level‌​‌ annotation scheme, moving from​​ general emotional descriptors (emotional​​​‌ activation), to basic (primary‌ emotions), and finally to‌​‌ fine-grained affective states (affective​​ phenomena). The analyses reveal​​​‌ a global reduction of‌ emotional expressiveness in HD‌​‌ compared to healthy controls.​​ Expressed emotions aligned with​​​‌ patients' psychiatric profiles (e.g.‌ anxious and nervous speech‌​‌ with depression; frustrated speech​​ with irritability and aggressivity),​​​‌ highlighting speech as a‌ promising marker for emotional‌​‌ and psychiatric symptoms, paving​​ the way for remote​​​‌ monitoring and personalised care‌ in HD. This PhD‌​‌ work is the result​​ of a collaboration between​​​‌ Inria, the Ecole Normale‌ Supérieure, and Inserm, and‌​‌ is funded by Université​​ Paris Cité (ED 474).​​​‌

Through several initiatives (BPI-funded‌ projects ONCOLAB and FG4H,‌​‌ collaboration with AP-HP, and​​ the INRIA AOC action),​​​‌ we continued to investigate‌ the use of NLP‌​‌ techniques, particularly LLMs, applied​​ to medical documents. This​​​‌ domain presents specific challenges:‌ clinical data (Electronic Health‌​‌ Records, EHRs) are highly​​ sensitive, difficult to access,​​​‌ and especially scarce in‌ French. Conversely, medical knowledge‌​‌ is more readily available​​ through ontologies, scientific publications,​​​‌ and guidelines. To address‌ these constraints, we are‌​‌ progressively developing a new​​ methodological paradigm based on:​​​‌

1. the generation of realistic‌ but non-confidential fictitious clinical‌​‌ data;
2. the automatic annotation​​ of this synthetic data​​​‌ using LLMs;
3. the training‌ of Small Language Models‌​‌ (SLMs) that can be​​ securely deployed on private​​​‌ infrastructures.

As part of‌ this effort, Rian Touchent‌​‌ finalised Biomed-Enriched, a two-step​​ pipeline for paragraph-level annotation​​​‌ of large public biomedical‌ corpora such as PubMed‌​‌ and ISTEX. In the​​​‌ first step, Llama-3.1-70B annotates​ a subset of PubMed​‌ paragraphs by type, domain,​​ and educational quality. These​​​‌ annotations are then used​ to train an SLM​‌ (XML-RoBERTa), which subsequently annotates​​ the full PubMed corpus.​​​‌ This enriched annotation enables​ fine-grained selection of pre-training​‌ datasets, for example targeting​​ high-quality clinical case descriptions.​​​‌ Experiments demonstrate that an​ encoder-based SLM trained on​‌ these curated subsets achieves​​ performance comparable to BioClinical-ModernBERT​​​‌ on 11 clinical benchmarks,​ while requiring 2.5 times​‌ less pre-training data, highlighting​​ the value of informed​​​‌ data selection.

Data selection​ techniques were also used​‌ by Rian Touchent to​​ train a new version​​​‌ of CamemBERT-bio (v2) 127​, 128 and to​‌ develop ModernCamemBERT-bio, a French​​ biomedical encoder based on​​​‌ ModernBERT/ModernCamemBERT with an extended​ context window (up to​‌ 8,192 tokens). Such extended​​ context capacity is particularly​​​‌ relevant for processing long​ clinical documents or full​‌ patient histories. Optimised GLiNER​​ variants of these models​​​‌ are also under development​ for information extraction tasks​‌ (ONCOLAB context) and ICD-10​​ coding (collaboration with AP-HP​​​‌ through the co-supervision of​ Anh Thu's M2 internship).​‌

For experiments on ICD-10​​ coding, models were trained​​​‌ on several thousand synthetic​ hospital reports generated by​‌ an LLM (Mistral) using​​ clinical scenarios and detailed​​​‌ medical guidelines. In parallel,​ Rian Touchent explored the​‌ reformulation of publicly available​​ PubMed clinical cases into​​​‌ standard hospital report formats.​ The original case provides​‌ the narrative structure and​​ essential medical content, which​​​‌ the model then adapts​ into the style of​‌ clinical notes.

Another line​​ of research concerns the​​​‌ generation of synthetic facsimile​ clinical documents, conducted as​‌ part of Simon Méoni​​ 's PhD work. The​​​‌ approach consists in extracting​ non-confidential UMLS concepts from​‌ real documents and using​​ them as inputs to​​​‌ train a generator model.​ Reinforcement learning is employed​‌ to optimise the similarity​​ between facsimile and real​​​‌ documents, thereby improving realism​ without compromising confidentiality. A​‌ complete workflow is now​​ available for producing such​​​‌ facsimile documents, particularly on​ the English MIMIC dataset,​‌ and was further validated​​ through the co-supervision of​​​‌ Tripodi's AP-HP internship. The​ objectives included generating French​‌ facsimiles and using them​​ to train multi-label ICD-10​​​‌ classifiers. Due to French​ regulatory constraints, the methodology​‌ prioritised synthetic documents authored​​ by clinicians rather than​​​‌ real patient data, thereby​ minimising re-identification risks. A​‌ significant part of Simon​​ Méoni 's work in​​​‌ 2025 was dedicated to​ evaluating the privacy risks​‌ associated with facsimile generation.​​ Although facsimiles contain no​​​‌ explicit identifiers, re-identification could​ still arise from unique​‌ sequences of UMLS concepts.​​ Several privacy metrics were​​​‌ identified and tested, leading​ to the development of​‌ mitigation procedures designed to​​ reduce potential confidentiality leaks​​​‌ while preserving the utility​ of the generated data​‌ for downstream tasks.

8.7.2​​ Patents

Our research on​​​‌ NLP for patents, carried​ out in collaboration with​‌ Questel (formerly with qatent,​​ acquired by Questel in​​​‌ 2024) in the context​ of You Zuo 's​‌ CIFRE PhD thesis (industrial​​ supervisor: Kim Gerdes), focuses​​​‌ on developing representation learning​ methods that respect the​‌ unique structure, length, and​​ technical density of patent​​ documents, with a particular​​​‌ emphasis on scalable and‌ annotation-light retrieval. In 2025,‌​‌ our first line of​​ work focused how self-supervised​​​‌ contrastive learning can be‌ adapted to the patent‌​‌ domain. We showed that​​ standard SimCSE-style dropout augmentation​​​‌ leads to overly uniform‌ embeddings for patents, due‌​‌ to their length and​​ internal redundancy. To address​​​‌ this, we introduced section-based‌ contrastive learning, where different‌​‌ patent sections (e.g. abstract,​​ claims, background) are treated​​​‌ as complementary semantic views.‌ This approach exploits patents'‌​‌ intrinsic discourse structure to​​ learn embeddings that preserve​​​‌ both global semantic coherence‌ and local continuity, achieving‌​‌ state-of-the-art performance in prior-art​​ retrieval and classification without​​​‌ relying on citation or‌ IPC (International Patent Classification)‌​‌ supervision 91.

Complementing​​ dense representations, our second​​​‌ line of work explores‌ geometry-aware sparse representations for‌​‌ patent retrieval. We initiated​​ the development of an​​​‌ unsupervised discretisation of dense‌ embedding spaces using overlapping‌​‌ hyperspherical coverings, producing learned​​ sparse vectors that capture​​​‌ both exact terminology and‌ semantic neighborhoods. These representations‌​‌ are directly compatible with​​ inverted indexes and significantly​​​‌ outperform traditional lexical and‌ clustering-based sparse baselines, while‌​‌ combining effectively with dense​​ retrievers. Together, these contributions​​​‌ advance robust, interpretable, and‌ scalable retrieval methods tailored‌​‌ to complex patent data.​​ We intend to publish​​​‌ on this research in‌ 2026.

8.8.1 French OLDI​

In collaboration with the​‌ Tralalam project, Malik Marmonier​​ , supervised by Rachel​​​‌ Bawden and Benoît Sagot​ , developed the first​‌ French partition of the​​ OLDI Seed Corpus 111​​​‌, our submission to​ the WMT 2025 Open​‌ Language Data Initiative (OLDI)​​ shared task 43.​​​‌ Its creation process involved​ using multiple MT systems​‌ and a custom-built interface​​ for post-editing by qualified​​​‌ native speakers. The source​ data presented unique translation​‌ challenges, as it combines​​ highly technical, encyclopedic terminology​​​‌ with the stylistic irregularities​ characteristic of user-generated content​‌ taken from Wikipedia. This​​ French corpus is not​​​‌ an end in itself,​ but is intended as​‌ a crucial pivot resource​​ to facilitate the collection​​​‌ of parallel corpora for​ the under-resourced regional languages​‌ of France, as discussed​​ previously in Section 8.3.4​​​‌.

8.8.2 Occitan Dialects​

Our work on Occitan​‌ corpus development, in the​​ context of Oriane Nédey​​​‌ 's PhD thesis supervised​ by Benoît Sagot ,​‌ Rachel Bawden and Thibault​​ Clérice , addresses the​​​‌ dual challenge of data​ scarcity and fine-grained linguistic​‌ variation across the Occitan​​ dialect continuum. In 2025,​​​‌ a first line of​ research focused on building​‌ ethically sound, linguistically rich​​ resources from UGC. We​​​‌ designed and documented a​ full pipeline for data​‌ collection, cleaning, and anonymisation​​ of forum data, ensuring​​​‌ compliance with ethical standards​ while preserving crucial sociolinguistic​‌ information 66. This​​ methodology underpins ForumOccitania, a​​​‌ large-scale corpus of online​ discussions that combines textual​‌ data with self-declared metadata​​ on dialect, geography, age,​​​‌ and speaker profile 84​. Beyond corpus construction,​‌ we carried out quantitative,​​ qualitative and exploratory computational​​​‌ analyses, showing that the​ corpus is predominantly written​‌ in Occitan using classical​​ orthography, largely by young​​​‌ neo-speakers, and that it​ exhibits clear dialectal signals​‌ from major varieties such​​ as Lemosin, Lengadocian, Gascon,​​​‌ and Provençau.

Complementing forum​ data, we developed OcWikiDialects,​‌ a large encyclopedic corpus​​ drawn from Occitan Wikipedia,​​​‌ enriched with explicit dialect​ labels covering eight major​‌ varieties and two transitional​​ ones. With millions of​​​‌ tokens and multiple textual​ representations and segmentations, this​‌ resource broadens the empirical​​ basis for studying written​​​‌ Occitan variation. Together, these​ corpora provide a coherent,​‌ methodologically grounded infrastructure for​​ sociolinguistic and NLP research​​​‌ on Occitan and its​ varieties.

8.8.3 Picard

In​‌ 2025, we continued the​​ collaborations established with the​​ Agence régionale de la​​​‌ Langue Picarde (ARLP) and‌ the Université de Picardie‌​‌ Jules Verne (UPJV). UPJV​​ provided the content of​​​‌ its textual database of‌ documents in Picard, PICARTEXT‌​‌, to COLaF. It​​ constitutes the largest dataset​​​‌ currently available for the‌ language, spanning from the‌​‌ 18th to the 21st​​ centuries and comprising 6​​​‌ million tokens. Work is‌ underway to transform this‌​‌ plain-text database into the​​ COLaF XML-TEI format, making​​​‌ it structured and searchable.‌ At the same time,‌​‌ UPJV is annotating the​​ dataset with linguistic information​​​‌ in order to train‌ language annotation models (e.g.‌​‌ lemmatisers), using the Pyrrha​​ application (a tool for​​​‌ post-correction of lemmatised and‌ morphosyntactic tagged corpora) on‌​‌ the instance provided by​​ COLaF on Huma-Num server.​​​‌

In parallel, a digitisation‌ campaign organised by Juliette‌​‌ Janès was carried out​​ at ARLP to recover​​​‌ analogue documents from their‌ archives, namely the Picard‌​‌ literature competitions from 2005​​ to 2020 (approximately 40​​​‌ documents per year). These‌ materials are currently being‌​‌ processed in XML-TEI using​​ the LADaS and ATR​​​‌ pipeline under development.

8.8.4‌ French-Based Creoles

Our contributions‌​‌ to French-based Creole studies​​ combine computational linguistics with​​​‌ historical linguistics to reassess‌ the development of French-based‌​‌ Creole languages through large-scale​​ textual evidence. This work​​​‌ is carried out primarily‌ by Rasul Jasir Dent‌​‌ in the context of​​ his PhD, supervised by​​​‌ Benoît Sagot , Thibault‌ Clérice and Pedro Ortiz‌​‌ (Common Crawl). A key​​ challenge in this regard​​​‌ is the creation of‌ new digital corpora designed‌​‌ to capture contact-induced variation​​ across time and space.​​​‌ In 2025, building on‌ the Molyé corpus, which‌​‌ brings together early Creole​​ attestations and European contact​​​‌ varieties over four centuries‌ 100, we have‌​‌ worked on corpus construction​​ methods tailored to sparsely​​​‌ documented languages. By reframing‌ language identification as a‌​‌ data-mining task rather than​​ full multi-class classification, we​​​‌ were able to rapidly‌ assemble broad and representative‌​‌ corpora for several French-based​​ Creoles, prioritising coverage over​​​‌ exhaustive document processing 74‌.

These resources underpin‌​‌ our historical investigations into​​ the relationships among Atlantic​​​‌ and Indian Ocean Creoles.‌ Based on our corpora,‌​‌ we provide evidence for​​ a nineteenth-century transoceanic continuum​​​‌ ranging from normative French‌ to basilectal Creole, with‌​‌ intermediate “Para-Creole” varieties showing​​ feature-level diffusion across regions​​​‌ 62. This corpus-driven‌ perspective also informs our‌​‌ re-evaluation of French-based pidgins.​​ A comparative analysis of​​​‌ Français Tirailleur and Tây‌ Bôi, supported by newly‌​‌ analysed textual data, suggests​​ that both were shaped​​​‌ by top-down pedagogical practices‌ rather than emerging solely‌​‌ from spontaneous communicative necessity​​ 61.

Finally, our​​​‌ corpus-building efforts have extended‌ to contemporary NLP infrastructure‌​‌ through shared evaluation campaigns,​​ resulting in new public​​​‌ datasets for multiple Creole‌ languages and fostering collaboration‌​‌ around their digital documentation​​ 48.

8.8.5 Language​​​‌ Identification

In 2025, again‌ in the context of‌​‌ Rasul Jasir Dent 's​​ PhD work, we redefined​​​‌ the language identification task‌ from the perspective of‌​‌ corpus creation for less​​ commonly written languages: rather​​​‌ than treating language identification‌ as a standard multi-class‌​‌ classification problem, we argued​​​‌ that it is more​ productively framed as a​‌ targeted data mining task,​​ in which only a​​​‌ small fraction of available​ documents are expected to​‌ be relevant 74,​​ 31. Building on​​​‌ this reframing, we designed​ an efficient pipeline that​‌ prioritises the rapid exclusion​​ of irrelevant material, thereby​​​‌ minimising computational and human​ resources spent on uninformative​‌ data, while making it​​ possible to find data​​​‌ for very scarcily attested​ language varieties (a “needle-in-a-haystack”​‌ scenario). Our approach makes​​ it possible to filter​​​‌ massive collections within a​ few hours, supporting the​‌ scalable construction of new​​ digital corpora 31.​​​‌ Although we primarily applied​ the methodology to French-based​‌ Creole languages (see above),​​ our proposed framework is​​​‌ generic and applicable to​ language identification scenarios characterised​‌ by extreme class imbalance,​​ and we carried out​​​‌ preliminary experiments to minority​ regional languages of France.​‌

8.8.6 Data Encoding

We​​ resumed our joint work​​​‌ on methods for encoding​ and representing linguistic diversity​‌ in text corpora, especially​​ for under-resourced, non-standard and​​​‌ historically situated varieties. In​ 2025 we developed a​‌ rigorous TEI-based schema that​​ captures fine-grained metadata for​​​‌ language variants across temporal,​ geographic and sociolinguistic dimensions,​‌ addressing a notable gap​​ in existing standards such​​​‌ as BCP47 by incorporating​ richer identifiers via resources​‌ like Glottolog 37.​​ This schema, whose main​​​‌ developer is Juliette Janès​ , plays a central​‌ role in COLaF, as​​ our aim is to​​​‌ apply it to all​ corpora distributed in this​‌ context.

Complementing this structural​​ work, we explored the​​​‌ diachronic and contact-influenced nature​ of language variation, exemplified​‌ by work on French-based​​ Creole (see above), which​​​‌ uses digitally compiled textual​ evidence to reconceptualise these​‌ varieties within networked historical​​ contexts rather than strict​​​‌ genealogical models 64.​

8.8.7 The Parallel Corpus​‌ of the Parable of​​ the Prodigal Son

In​​​‌ order to cover the​ diversity of the regional​‌ language varieties spoken in​​ metropolitan France, in 2025​​​‌ we started to build​ a parallel corpus of​‌ the Parable of the​​ Prodigal Son, a collection​​​‌ of translations of this​ parable produced for national​‌ linguistic surveys in the​​ 19th century. We collected​​​‌ the data with the​ help of two external​‌ researchers, Sven Ködel and​​ Alexandre Génadot, structured the​​​‌ data and language metadata​ in XML-TEI files (See​‌ Section 8.8.6), aligned​​ the versions and established​​​‌ a comparison between these​ data and the Atlas​‌ Linguistique de France ones​​ by projecting lexical data​​​‌ on georeferenced maps. The​ work is still in​‌ progress (increase of data,​​ improvement of the automatisation​​​‌ of the full pipeline)​ and a publication is​‌ upcoming.

8.9.1 Epistemic approach to‌​‌ ATR

The year 2025​​ was marked by a​​​‌ strong epistemological engagement with‌ the meaning, scope, and‌​‌ implications of transcription and​​ automatic transcription in the​​​‌ humanities. This work resulted‌ in several publications, including‌​‌ the preparation of a​​ collective book and multiple​​​‌ research articles.

The book‌ Apprendre à lire aux‌​‌ machines80, currently​​ available as a preprint​​​‌ and co-edited by Thibault‌ Clérice , Alix Chagué‌​‌ , and Ariane Pinche​​ (CNRS), constitutes the first​​​‌ French-language volume devoted entirely‌ to ATR for the‌​‌ humanities. It brings together​​ contributions on seven writing​​​‌ systems (Latin, Greek, Armenian,‌ Arabic, Chinese, Bengali, and‌​‌ Urdu), alongside four transversal​​ chapters addressing broader issues​​​‌ such as ATR infrastructures,‌ methodological choices, and the‌​‌ historical development of the​​ field. This publication was​​​‌ complemented by further studies‌ examining the impact of‌​‌ ATR on humanities research​​ 16 as well as​​​‌ its current limitations in‌ relation to model architectures‌​‌ and available datasets 20​​, 65.

The​​​‌ year concluded with a‌ preprint by Benjamin Kiessling‌​‌ offering a fine-grained analysis​​ of transcription guideline variability​​​‌ in the humanities, the‌ tension between traditional scholarly‌​‌ transcription and transcription optimised​​ for ATR, and the​​​‌ implications of these choices‌ for model generalisation across‌​‌ scripts 81.

8.9.2​​ Improvement and Consolidation of​​​‌ Existing Results

In contrast‌ to previous years focused‌​‌ on rapid methodological advances,​​ 2025 was primarily dedicated​​​‌ to the stabilisation, consolidation,‌ and refinement of existing‌​‌ tools, models, and standards.​​ This included the publication​​​‌ of a reference paper‌ on Kraken V5 38‌​‌, the ATR engine​​ to whose development Inria​​​‌ has significantly contributed, as‌ well as improvements in‌​‌ dataset documentation through enhanced​​ schemas for HTR-United 21​​​‌ and a revised version‌ of the LADaS annotation‌​‌ guidelines 79.

Additional​​ work addressed the transcription​​​‌ of modern scripts in‌ a European context over‌​‌ the last five centuries​​ 63. For medieval​​​‌ material, data production continued‌ within the HTRogène project,‌​‌ although without new publications,​​ apart from the release​​​‌ of updated models (CATMuS‌ Medieval 1.6.0, 97).‌​‌ Finally, ongoing but as​​ yet unpublished work focused​​​‌ on improving layout analysis‌ within Kraken and eScriptorium;‌​‌ tangible results from these​​ efforts are expected in​​​‌ 2026.

8.9.3 New Challenges‌ in ATR

Beyond consolidation,‌​‌ four emerging challenges were​​ explored in 2025, although​​​‌ they have not yet‌ resulted in substantial published‌​‌ output.

The first challenge​​ arose within the Back​​​‌ In Time project and‌ concerns the automatic transcription‌​‌ of enciphered texts composed​​ of unique or highly​​​‌ idiosyncratic glyphs. Addressing this‌ problem required pushing ATR‌​‌ methodologies beyond standard assumptions​​ about scripts and Unicode-available​​​‌ signs. This work led‌ to the development of‌​‌ Glyphea, an extension​​ for eScriptorium specifically designed​​​‌ to handle complex, non-standard‌ scripts and enciphered writing‌​‌ systems, as well as​​​‌ to broader improvements to​ eScriptorium in this direction.​‌

The second challenge concerns​​ the development of more​​​‌ generalisable and language-independent ATR​ models. In this context,​‌ ALAMAnaCH is organising the​​ ICDAR 2026 Competition on​​​‌ Multilingual Medieval Handwriting Recognition,​ coordinated by Benjamin Kiessling​‌ and Thibault Clérice ,​​ with the goal of​​​‌ fostering methodological advances in​ cross-lingual and cross-script generalisation.​‌

The third challenge focuses​​ on the automatic transcription​​​‌ and structural formalisation of​ edited Ancient Greek texts.​‌ Conducted within the Corpus​​ Liberatum Linguae Graecae project​​​‌ (Thibault Clérice ,​ Nicolas Angleraud , Antonia​‌ Karamolegkou , Benoît Sagot​​ ), this work evaluates​​​‌ the ability of both​ ATR systems and Visual​‌ LLMs (VLLMs) to recover​​ structural information and Ancient​​​‌ Greek text from printed​ scholarly editions 57.​‌ In 2025, foundational work​​ was completed through the​​​‌ creation of a synthetic​ dataset, with a publication​‌ planned for early 2026.​​ A fourth challenge is​​​‌ being addressed within the​ PhD project of Hugo​‌ Scheithauer . His research​​ focuses on Optical Music​​​‌ Recognition (OMR) for historical​ documents and will begin​‌ to yield published results​​ in 2026. This work​​​‌ also includes the release​ of a first dataset​‌ dedicated to the transcription​​ of 19th-century​​​‌ musical manuscripts. The PhD​ is conducted under the​‌ supervision of Thibault Clérice​​ , Gonzalo Romero-García (Epita),​​​‌ and Laurent Romary.

8.9.4​ Application of ATR at​‌ scale

Finally, 2025 represented​​ a turning point in​​​‌ the large-scale application of​ ATR expertise through the​‌ creation of CoMMA (Corpus​​ of Multilingual Medieval Archives).​​​‌ This corpus comprises approximately​ 33,000 manuscripts automatically transcribed​‌ using CATMuS models over​​ a four-month period, in​​​‌ collaboration with EquipEx Biblissima+​ and three humanities researchers​‌ 73.

The resulting​​ corpus, amounting to roughly​​​‌ 3.3 billion tokens of​ Latin and Old French​‌ texts spanning the 9​​th to the 16​​​‌th centuries, is accessible​ through a dedicated user​‌ interface, TexTile (see the​​ CoMMA web site for​​​‌ more details). TexTile relies​ on the Distributed Text​‌ Services protocol 59,​​ 60 to enable scalable​​​‌ access and exploration. Work​ is currently underway to​‌ extend CoMMA to additional​​ digitisation repositories and to​​​‌ other languages, notably from​ the Iberian Peninsula, in​‌ 2026.

8.11.1 Using Interbrain Synchrony​​​‌ and Rapport Building to‌ Understand Productive Peer Collaboration‌​‌

As part of our​​ large-scale research programme on​​​‌ collaboration, our interdisciplinary team‌ continued and extended its‌​‌ investigation of dyadic collaboration​​​‌ in children aged 5​ to 12. The overarching​‌ objective of this work​​ is to better understand​​​‌ the mechanisms underlying successful​ collaboration by examining the​‌ relationships between interpersonal rapport,​​ collaborative performance, and inter-brain​​​‌ synchrony (IBS), measured using​ functional near-infrared spectroscopy (fNIRS)​‌ hyperscanning. Building on the​​ experimental framework established previously,​​​‌ the study focuses on​ how children's language and​‌ non-verbal behaviours relate to​​ both learning outcomes and​​​‌ neural coupling between interacting​ peers during collaborative problem-solving​‌ tasks. In 2025, the​​ project introduced a focus​​​‌ on transactivity, a key​ dimension of collaboration defined​‌ as the extent to​​ which participants take up,​​​‌ respond to, and build​ upon one another's ideas​‌ during interaction. Within this​​ framework, the team conducted​​​‌ a longitudinal, multimodal data​ collection in Parisian primary​‌ schools. Data were collected​​ from 13 dyads of​​​‌ children over a four-week​ period using fNIRS hyperscanning.​‌ In addition, 8 dyads​​ from another age group​​​‌ were recorded using the​ same protocol, enabling replication​‌ and comparison across three​​ distinct age groups. This​​​‌ expanded dataset enables us​ to advance the preliminary​‌ investigation of the relationship​​ between transactivity, learning gains,​​​‌ and inter-brain synchrony.

8.11.2​ Son of Sara: Developing​‌ a new LLM-based Embodied​​ Conversational Agent

As part​​​‌ of our ongoing work​ on socially capable conversational​‌ agents, we continued the​​ development of Son of​​​‌ SARA, an embodied conversational​ agent designed to support​‌ natural and effective interaction​​ with human users. The​​​‌ project aims to equip​ dialogue agents with both​‌ verbal and non-verbal interactional​​ skills that are essential​​​‌ for collaborative communication. In​ 2025, the work focused​‌ on extending the agent​​ beyond speech-only interaction by​​​‌ adding a visual and​ embodied dimension to the​‌ conversation. Building on an​​ existing conversational framework enabling​​​‌ spoken dialogue, we developed​ the agent's virtual body​‌ and implemented the software​​ infrastructure required to synchronise​​​‌ non-verbal behaviours with speech.​ This included the design​‌ and implementation of preliminary​​ rule-based models for the​​​‌ generation of facial expressions​ and gestures, allowing the​‌ agent to produce visible​​ communicative signals aligned with​​​‌ its speech turns. In​ parallel, the project advanced​‌ the agent's conversational dynamics​​ by improving its turn-taking​​​‌ capabilities. A predictive model​ based on voice activity​‌ was integrated to enable​​ the agent to anticipate​​​‌ turn transitions during interaction.​ This contributes to more​‌ interactive and rhythmically natural​​ exchanges, bringing agent-human conversations​​​‌ closer to the temporal​ structure of human-human dialogue.​‌ Together, these developments mark​​ a significant step toward​​​‌ the fully embodied nature​ of the Son of​‌ SARA conversational agent. They​​ provide a solid foundation​​​‌ for future work on​ data-driven, real-time gesture generation​‌ conditioned on vocal features,​​ as well as on​​​‌ further refinements of turn-taking​ models, including adaptation to​‌ additional languages.

8.11.3 Conversational​​ Grounding in Dialogue Systems​​​‌

As part of our​ research on socially competent​‌ dialogue systems, we continued​​ our investigation of conversational​​​‌ grounding, a fundamental mechanism​ through which interlocutors establish​‌ and maintain mutual understanding​​ over the course of​​​‌ an interaction. This work​ focuses on the ability​‌ of LLMs and LLM-based​​ agents to represent, update,​​ and exploit shared knowledge​​​‌ during dialogue. In 2025,‌ the research shifted from‌​‌ grounding limited to immediate​​ conversational context toward the​​​‌ modelling of long-term common‌ ground, that is, information‌​‌ accumulated and reused across​​ extended interactions. To support​​​‌ this objective, we introduced‌ Indiref, a new benchmark‌​‌ designed to assess the​​ use of grounded information.​​​‌ This gave us a‌ way to test different‌​‌ representation techniques for common​​ ground including textual and​​​‌ visual representations. Our evaluation‌ revealed that standard representation‌​‌ techniques performed poorly on​​ this benchmark. In response,​​​‌ we developed a pipeline‌ to generate synthetic situated‌​‌ conversations, which were then​​ used to train LLMs​​​‌ via reinforcement learning. We‌ successfully demonstrated that this‌​‌ approach significantly enhances grounding​​ performance, allowing us to​​​‌ isolate the key requirements‌ for establishing robust long-term‌​‌ capabilities in conversational agents.​​ All of this work​​​‌ is described in detail‌ in an article to‌​‌ appear in 2026; a​​ preprint is already available​​​‌ 83.

8.11.4 Exploring‌ Interpersonality: Multimodal Personality Cues‌​‌ in Embodied Conversational Agents​​

As part of our​​​‌ research on personality expression‌ in embodied conversational agents,‌​‌ we continued our investigation​​ of interpersonality, a framework​​​‌ that captures how the‌ multimodal manifestation of personality‌​‌ emerges from interaction and​​ is shaped by the​​​‌ interlocutor, the communicative context,‌ and broader situational factors.‌​‌ In 2025, we successfully​​ obtained ethical approval from​​​‌ COERLE, enabling the launch‌ of a new experimental‌​‌ study involving the collection​​ of a corpus of​​​‌ quadrads (groups of four‌ participants). The experimental design‌​‌ was substantially revised in​​ close collaboration with the​​​‌ KETI team, with the‌ objective of enabling optimal‌​‌ observation of gestures and​​ embodied behaviors, and ensuring​​​‌ the creation of a‌ replicable corpus. This design‌​‌ choice lays the groundwork​​ for future cross-cultural analyses​​​‌ of interpersonality and multimodal‌ interaction dynamics. This work‌​‌ establishes the methodological and​​ empirical foundations for future​​​‌ analyses of interpersonality and‌ multimodal interaction dynamics, and‌​‌ supports the long-term objective​​ of understanding how personality​​​‌ is jointly constructed through‌ social interaction across cultural‌​‌ contexts.

LightON

• Partner type:​
  Start-up
• Leader for ALMAnaCH:​‌
  Djamé Seddah.
• Dates:
  22​​ Sept 2020–present
• Description:
  LightON​​​‌ used to build Optical​ Processor Units, a specialised​‌ line of processor able​​ to outperform GPUs on​​​‌ certains tasks and since​ the release of the​‌ Pagnol models in 2021,​​ with the participation of​​​‌ Djamé Seddah, has been​ deploying LLMs in various​‌ industrial contexts ever since.​​ This then informal collaboration​​​‌ has led to a​ successful BPI "communs numériques"​‌ project named Scribe where​​ both LightOn and Almanach​​ are co-PIs with two​​​‌ CIFRE PhDs starting in‌ February 2026, both co-supervised‌​‌ by Djamé Seddah and​​ Amélie Chatelain and Iacopo​​​‌ Poli (Lighton). They will‌ in post-training strategies (Reinforcement‌​‌ learning, Reasonning models, synthetic​​ data generation, etc.).

zaion​​​‌

Participants: Chloé Clavel,‌ Lorraine Vanel.

• Partner‌​‌ type:
  Company
• Leader for​​ ALMAnaCH:
  Chloé Clavel.
• Dates:​​​‌
  1 Feb 2022–1 Mar‌ 2025
• Funding received:
  €16,000‌​‌
• Description:
  CIFRE PhD thesis​​ between Telecom-Paris and Zaion​​​‌ in order to develop‌ conversational systems integrating socio-emotional‌​‌ strategies in an explicit​​ way. The CIFRE contract​​​‌ being with TelecomParis, it‌ is mentioned in this‌​‌ section rather than in​​ the previous one.

10.1.1 Associate‌​‌ Teams in the framework​​ of an Inria International​​​‌ Lab or in the‌ framework of an Inria‌​‌ International Program

10.1.2​​​‌ Participation in other International‌ Programs

10.2.1 Visits of international​​​‌ scientists

10.3.1​​ Horizon Europe

10.4.1 ANR​​​‌

10.4.2 Competitivity Clusters and​ Thematic Institutes

10.4.3‌ Other National Initiatives

Domaine​​​‌ de recherche et d'innovation‌ majeurs (DIM) AI4IDF

Participants:‌​‌ Chloé Clavel, Khaled​​ Benaida.

• Duration:
  1​​​‌ Sept 2021–1 Sept 2026.‌
• PI:
  Chloé Clavel
• Partners:‌​‌
  • PRAIRIE
  • DataIA
  • Hi!Paris
  • SCAI​​
• Summary:
  AI4IDF aims to​​​‌ deepen knowledge in AI‌ while keeping the human‌​‌ being at the center​​ of its concerns. The​​​‌ Paris Region must play‌ a major role in‌​‌ this future sector, thanks​​ to its scientific excellence​​​‌ and its incomparable ecosystem.‌

Domaine de recherche et‌​‌ d'innovation majeurs (DIM) Patrimoines​​ matériels - innovation, expérimentation​​​‌ et résilience

Participants: Alix‌ Chagué, Thibault Clérice‌​‌.

• Duration:
  1 Jan​​ 2022–31 Dec 2026.
• Coordinator​​​‌ for ALMAnaCH:
  Laurent Romary‌ .
• Summary:
  The DIM‌​‌ Patrimoines matériels - innovation,​​ expérimentation et résilience (PAMIR)​​​‌ aims to bring out‌ new forms of social,‌​‌ environmental and economic development​​ by connecting museums, companies,​​​‌ the Ile-de-France ecosystem of‌ creating and crafts, universities,‌​‌ and laboratories around questions​​ of fundamental and applied​​​‌ research on heritage collections‌ and issues.

11.1.1 Scientific events: organisation​​​‌

11.1.2 Scientific‌​‌ events: selection

11.1.3 Journal

11.1.4 Invited talks​‌

• Alix Chagué :
  • Institute​​ of Advanced Studies, Princeton,​​​‌ Etats-Unis (13 Jun 2025):​ “HTR-United schema for dataset​‌ descriptions”.
• Chloé Clavel :​​
  • Saint-Denis, France (2 Apr​​​‌ 2025): “Best Practices”. Symposium​ Européen du Numérique et​‌ de l'IA Responsables (Région​​ Ile de France)
  • AI​​​‌ Grid Spring School (7​ Apr 2025): “Generative AI​‌ and Social Interactions”. Website​​
  • TALEP team, Aix-Marseille University​​​‌ (25 Sept 2025): “Computational​ Models of Socio-emotional Interactions​‌ in the Era of​​ LLMs - the Challenges​​​‌ of Transparency”.
  • “Understanding social​ interactions in the era​‌ of LLMs: The challenges​​ of transparency”, LATTICE, ENS​​​‌ (14 Oct 2025), Keynote​ - European Conference on​‌ Artificial Intelligence, ECAI 2025,​​ LUHME workshop Language Understanding​​​‌ in the Human-Machine Era,​ Bologne, Italie (26 Nov​‌ 2025), and DIC/ISC/CRIA Seminar​​ in Cognitive Informatics, Université​​​‌ du Québec à Montréal​ (UQÀM) (27 Nov 2025).​‌
  • Webinar Matching of PEPR​​ Ensemble (18 Dec 2025):​​​‌ “NLP-Driven Models of Collaboration​ in Human-Human and Human-Agent​‌ Interactions”.
• Éric Villemonte De​​ La Clergerie :
  • Webinar​​​‌ IIAS PROMESS Staff Week​ (online) (18 Nov 2025):​‌ “Towards more open medical​​ datasets ?”.
  • Sofcot -​​​‌ 99ème Congrès de la​ Société Française de Chirurgie​‌ Orthopédique & Traumatologique, Paris​​ (12 Nov 2025): “Keynote​​​‌ IA : enjeux de​ fiabilité et de recevabilité​‌ des LLM”. Joint with​​ Jules Descamps.
• Célia Nouri​​​‌ :
  • Journée d'étude de​ l'Arcom (13 Nov 2025):​‌ “Contextualiser les interactions en​​ ligne : Des modèles​​​‌ pour détecter les interventions​ critiques et le langage​‌ abusif”. Website
  • Seminar of​​ the Digital Regulation Expertise​​​‌ Center (PEReN), Ministère de​ l'Économie et des Finances,​‌ Paris (1 Jul 2025):​​ “Detecting abusive language in​​​‌ online conversations with graph​ models”.
• Carlo Santagiustina :​‌
  • Brussels Institute for Advanced​​ Studies (BRIAS), Brussels, Belgium.​​​‌ (5 Dec 2025): “Representation,​ Political Discontent and e-Petitions:​‌ How economic ideology and​​ anti-elitism drive online petitioning​​​‌ in the EU”. “Resilient​ Democracy Online: Reimagining Civic​‌ Debate on Social Media”​​ workshop
• Élodie Étienne :​​​‌
  • ISIR, campus de Jussieu,​ Paris 5e (15 Dec​‌ 2025): “Étude des biais​​ de genre dans la​​​‌ collaboration multimodale humain–agent :​ entre observation et simulation”.​‌ Joint with Chloé Clavel​​ , Magalie Ochs. Journée​​​‌ “Intelligence Artificielle, Interactions Socio-Affectives​ et Éthique”
• Justine Cassell​‌ :
  • Paris, France (7​​ Feb 2025): “What neuroscience​​ can teach us about​​​‌ AI-child interaction”. AI impact‌ in Sciences and Humanities‌​‌ - AI Action Summit's​​ “AI, Science and Society”​​​‌ conference
  • Dublin, Ireland (4‌ Mar 2025): “AI as‌​‌ learning partner”. Moving from​​ AI Safety to AI​​​‌ Quality
  • “Growing up in‌ the Digital Age” Conference‌​‌ (4 Mar 2025): “GenAI​​ as learning partner not​​​‌ oracle”.
  • Kyoto, Japan (1‌ Oct 2025): “IA and‌​‌ education”. 22nd Annual Meeting​​ of Science Technology in​​​‌ Society forum (STS Forum)‌
  • Montreal, Canada (28 Nov‌​‌ 2025): “Overview of PRAIRIE”.​​ French AI Clusters
  • Montreal,​​​‌ Canada (28 Nov 2025):‌ “AI and Education”. Education‌​‌ and AI
  • University of​​ Oxford, London (1 Jan​​​‌ 2025): “Gesture Understanding and‌ Gesture Generation with new‌​‌ AI models”.
• Lucence Ing​​ :
  • CTHS, Orléans (17​​​‌ Apr 2025): “Le traitement‌ du lexique dans les‌​‌ premiers imprimés, entre conservation​​ et renouvellement : étude​​​‌ comparative des éditions princeps‌ du Lancelot et du‌​‌ Tristan en prose”. Session​​ “Transposer, actualiser et/ou refonder​​​‌ : textes et autorités‌ à l'épreuve de l'imprimé”‌​‌ au 149e congrès du​​ CTHS
  • Université de Padoue,​​​‌ Italie (22 May 2025):‌ “Élargir la portée d'un‌​‌ modèle d'annotation linguistique :​​ problèmes et perspectives pour​​​‌ le français médiéval”. Journée‌ “Testi, varietà linguistiche e‌​‌ trattamento automatico del 'francese​​ d'Italia”'
  • ENC, PSL, Paris​​​‌ (11 Jun 2025): “A‌ Study of Lexical Evolution‌​‌ in Medieval French: Digital​​ Approaches to Arthurian Romances”.​​​‌ Journée “Digital Approaches to‌ Pre-Modern Texts and Manuscripts”‌​‌
• Lydia Nishimwe :
  • Centrale​​ Nantes Alumni, Paris (25​​​‌ May 2025): “Table-ronde :‌ Les Centraliennes dans les‌​‌ métiers de la recherche”.​​ One of 5 panelists.​​​‌
• Rian Touchent :
  • Inist-CNRS‌ (16 Jun 2025): “L'utilisation‌​‌ de corpus textuels pour​​ l'entraînement des modèles de​​​‌ langage”. Journées Istex 2025‌
• Benoît Sagot :
  • Awayday‌​‌ (“journée au vert”) of​​ the Inria Lyon research​​​‌ center in Saint-Étienne (7‌ Jul 2025): “Apprendre les‌​‌ langues aux machines ?”.​​
• Djamé Seddah :
  • “Preventing​​​‌ Language Models Weaponization: What‌ if their Training Data‌​‌ were to be Compromised?”.​​ Social Science and Generative​​​‌ AI: Inquiries, Instruments, Consequences‌ (Conférence), Science Po (5‌​‌ Jun 2025) and Journées​​ INESIA, Inria, Paris (8​​​‌ Aug 2025).
  • CNIT, Paris‌ (9 Dec 2025): “Developing‌​‌ LLMs on academic budget,​​ are we building the​​​‌ next Maginot Line? a‌ Perspective from the GAPeron‌​‌ Experience”. OpenLLM Builder Days@FOST​​
• Thibault Clérice :
  • King's​​​‌ College of London, London,‌ United Kingdom (20 Jan‌​‌ 2025): “Building a cross-lingual​​ dataset from medieval manuscript​​​‌ text recognition, challenges and‌ outcomes of CATMuS”.
  • MSHS.‌​‌ Lille (13 May 2025):​​ “Collaboration et valorisation des​​​‌ codes et logiciels :‌ défis, retours d'expérience et‌​‌ perspectives”.
  • MSHS. Lille (14​​ Nov 2025): “Des données​​​‌ de projet aux données‌ de projets: exemples de‌​‌ réutilisations de données”.
  • Université​​ de Genève, Geneva, Switzerland​​​‌ (16 Dec 2025): “CoMMA‌ A Large-scale Corpus of‌​‌ Multilingual Medieval Archives”. Digital​​ humanities seminar.
• Wissam Antoun​​​‌ :
  • American University of‌ Beirut, Lebanon (25 Oct‌​‌ 2025): “AraBERT: An enabler​​ for Arabic NLP”.

11.1.5​​​‌ Scientific expertise

• Chloé Clavel‌ :
  • HCERES Reviewer for‌​‌ GIPSA-LAB.
  • Member of ANR​​ CE 38.
• Justine Cassell​​​‌ :
  • Reviewer for the‌ ERC-SYG-SH call.
  • Reviewer for‌​‌ ANR AAPG.
  • Reviewer for​​​‌ DAAD Postdoc-NeT-AI (Reviewed project​ proposal).
  • Member of the​‌ Conseil National du Numérique.​​
• Rachel Bawden :
  • Reviewer​​​‌ of the thematic committee​ for Genci projects (Reviewing​‌ projects for the allocation​​ of computational resources).
• Benoît​​​‌ Sagot :
  • Member of​ the scientific advisory board​‌ of ERIC CLARIN.
  • Member​​ of the Scientific Council​​​‌ of ARTE France.
  • Member​ of IA Action Summit​‌ steering committee (Paris 2025).​​
  • Member of Scientific Council​​​‌ of CNAM school on​ AI and digital sciences.​‌
  • Expert of Consultation by​​ a committee on AI-related​​​‌ legal matters put in​ place by CSPLA, the​‌ French Higher Council for​​ Literary and Artistic Property​​​‌ (17/11/2025).

11.1.6 Research administration​

• Rachel Bawden :
  • Member​‌ of the scientific board​​ of the Société de​​​‌ Linguistique de Paris (Co-administratrice).​
  • Member of the scientific​‌ board of EAMT executive​​ committee (Co-opted member).
• Benoît​​​‌ Sagot :
  • Member of​ the scientific board of​‌ Inria Paris's Comité des​​ Projets (Inria Paris research​​​‌ centre's Bureau du Comité​ des Projets).
  • Member of​‌ the scientific board of​​ the Société de Linguistique​​​‌ de Paris (Co-administrateur).

11.2.1 Teaching​​​‌

• Rachel Bawden :
  • Master's​ course (M2), Master “Mathématiques,​‌ Vision Apprentissage”, ENS Paris-Saclay,​​ France. CM: Speech and​​​‌ Language Processing, co-organised​ with Chloé Clavel ,​‌ Benoît Sagot , and​​ Emmanuel Dupoux. 3hrs.
• Justine​​​‌ Cassell :
  • Master's course​ (M2), Cerveau Social, Sorbonne​‌ Université, Paris, France. CM:​​ Metacognition dans l'interaction virtuelle​​​‌.
• Chloé Clavel :​
  • Master's course (M2), Artificial​‌ Intelligence & Advanced Visual​​ Computing Master, Polytechnique. CM:​​​‌ Speech emotion recognition, speech​ synthesis and conversational systems​‌. 6hrs.
  • Master's course​​ (M2), Mastère spécialisé en​​​‌ Intelligence Artificielle, Telecom-Paris. CM:​ Introduction to Natural Language​‌ Processing. 3hrs.
  • Master's​​ course (M2), Master “Mathématiques,​​​‌ Vision Apprentissage”, ENS Paris-Saclay,​ France. CM: Speech and​‌ Language Processing, co-organised​​ with Rachel Bawden ,​​​‌ Benoît Sagot , and​ Emmanuel Dupoux. 6hrs.
  • Master's​‌ course (M2), X Datascience,​​ Telecom-Paris. CM: Sentiment analysis​​​‌ and conversational systems.​ 6hrs.
• Thibault Clérice :​‌
  • Bachelor's Master's and Doctorate​​ level, TranscriboQuest Summer School,​​​‌ ENS Lyon. CM: HTR​ for Historical Documents,​‌ co-organised with Ariane Pinche.​​ 18hrs.
• Aina Garí Soler​​​‌ :
  • Master's course (M2),​ Science & Technology for​‌ Health (ST4Health), ESPCI, PSL.​​ CM: Artificial Intelligence Applied​​​‌ to Mental Health.​ 21hrs.
• Cecilia Graiff :​‌
  • Master's course (M2), Introduction​​ to data analysis with​​​‌ Python, Sciences Po Paris.​ CM: Introduction to data​‌ analysis with Python.​​ 12hrs.
  • Master's course (M1),​​​‌ Projet de statistique et​ science des données appliquées,​‌ ENSAE. Project supervision: Sociodemographic​​ debiasing influence on fake​​​‌ news detection, co-organised​ with Gabrielle Le Bellier​‌ . 10hrs.
• Lucence Ing​​ :
  • Master's course (M1),​​​‌ Master Humanités numériques, ENC,​ PSL. TD/TP: Structuration XML​‌ et XML/TEI. 10hrs.​​
  • Master's course (M1), Master​​​‌ Humanités numériques, Université de​ Rouen Normandie. TD/TP: Encodage​‌ des données : introduction​​. 15hrs.
  • Master's course​​​‌ (M1), Master Humanités numériques,​ Université de Rouen Normandie.​‌ CM: Introduction aux humanités​​ numériques. 14hrs.
  • CRBC/MSHB,​​​‌ Brest. Workshop on eScriptorium​ for HTR projects: Initiation​‌ à eScriptorium. 6hrs.​​
  • ENC, PSL, Paris. Part​​ of the workshop “Digital​​​‌ Approaches to Pre-Modern Texts‌ and Manuscripts”: A Multilingual‌​‌ Medieval Aligner: from Raw​​ Data to Aligned Witnesses​​​‌. 3hrs.
  • Biblissima, Sète.‌ Journées des clusters 5b‌​‌ et 7 Biblissima+: Aquilign,​​ un outil d'alignement multilingue​​​‌ pour corpus multilingues,‌ coorganised with Matthias Gille‌​‌ Levenson, Carolina Macedo. 3hrs.​​
• Gabrielle Le Bellier :​​​‌
  • Master's course (M1), Projet‌ de statistique et science‌​‌ des données appliquées, ENSAE.​​ Project supervision: Sociodemographic debiasing​​​‌ influence on fake news‌ detection, coorganised with‌​‌ Cecilia Graiff . 10hrs.​​
• Célia Nouri :
  • Master's​​​‌ course (M2), Master SCIA,‌ NLP track, EPITA, Paris.‌​‌ CM: Advanced Natural Language​​ Processing, co-organised with​​​‌ Francis Kulumba , Rian‌ Touchent . 46hrs.
  • Master's‌​‌ course (M2), Master Data​​ for Business, Kedge Bordeaux.​​​‌ CM: The Social Impacts‌ of Artificial Intelligence.‌​‌ 15hrs.
  • Bachelor's course (2nd​​ year bachelor), Collège universitaire,​​​‌ Sciences Po Paris. TD/TP:‌ Sciences et sociétés.‌​‌ 15hrs.
  • Master's course (M2),​​ Master Public Affairs, Track​​​‌ Digital and Technolgy, Sciences‌ Po Paris. TD/TP: Decoding‌​‌ Biases in Artificial Intelligence​​, co-organised with Jean-Philippe​​​‌ Cointet. 15hrs.
  • Master's course‌ (M2), Master “Mathématiques, Vision‌​‌ Apprentissage”, ENS Paris-Saclay, France.​​ Project supervision for the​​​‌ Speech and Language Processing‌ course. 4hrs.
• Benoît Sagot‌​‌ :
  • we are_, Paris.​​ IA training by We​​​‌ Are_ school for the‌ Casino Group Executive Board‌​‌ and for the Board​​ of Directors: Introduction au​​​‌ TAL et à l'IA‌ générative. 5hrs +‌​‌ 5hrs.
  • Servier headquarters, Suresnes,​​ France. Invited talk at​​​‌ the AI Day of‌ the Institut Servier de‌​‌ Médecine Translationnelle: Teaching languages​​ to machines. 1hr.​​​‌
  • Invited talk as part‌ of the “Cycle supérieur‌​‌ du numérique” from IGPDE:​​ Apprendre les langues aux​​​‌ machines — 2 ans‌ après. 1hr.
  • Master's‌​‌ course (M2), Master “Mathématiques,​​ Vision Apprentissage”, ENS Paris-Saclay,​​​‌ France. CM: Speech and‌ Language Processing, coorganised‌​‌ with Benoît Sagot ,​​ Chloé Clavel and Emmanuel​​​‌ Dupoux. 6hrs.
• Carlo Santagiustina‌ :
  • Moulin d'Andé, 65‌​‌ Rue du Moulin, Andé,​​ France. Invited to participate​​​‌ and contribute at the‌ “Université d'hiver du projet‌​‌ RésIn”: L'intelligence artificielle et​​ les sciences sociales Questions,​​​‌ méthodes et pratiques,‌ co-organised with SciencesPo médialab‌​‌ & Université Paris Cité.​​ 27hrs.
• Djamé Seddah :​​​‌
  • Master's course (Executive education),‌ Certificat Chef de projet‌​‌ IA, PSL-Dauphine. CM: Natural​​ Language Processing. 14hrs.​​​‌
  • Master's course (M2), Master‌ Linguistics, track Computational Linguistics,‌​‌ Université Paris Cité. CM:​​ Broadening NLP: Pretraining, instruct,​​​‌ alignment. 16hrs.
• Yi‌ Yu :
  • Master's course‌​‌ (M1), Machine learning for​​ Text Mining, Télécom-Paris. TD/TP:​​​‌ HMM, coorganised with‌ Matthieu Labeau and Maria‌​‌ Boritchev. 3hrs.
  • Bachelor's course,​​ Cycle Pluridisciplinaire d'Enseignement Supérieur​​​‌ 24/25, CPES. TD/TP: Cycle‌ Pluridisciplinaire d'Enseignement Supérieur 24/25‌​‌, coorganised with Maria​​ BORITCHEV. 7hrs.
  • Master's course​​​‌ (Mastères Spécialisés), IA717 Natural‌ Language Processing 2025/2026, Télécom-Paris.‌​‌ Project supervision: IA717 Natural​​ Language Processing 2025/2026,​​​‌ coorganised with Matthieu Labeau‌ and Maria Boritchev. 18hrs.‌​‌
  • Master's course (Mastères Spécialisés),​​ CSC_5DS25_TP Natural Language Processing​​​‌ and Sentiment Analysis 2025/2026,‌ Institut Polytechnique de Paris.‌​‌ TD/TP: Natural Language Processing​​, coorganised with Matthieu​​​‌ Labeau. 7hrs.
• You Zuo‌ :
  • Master's course (M1),‌​‌ Master pluriTAL, INALCO, France.​​​‌ CM: NLP in English​. 52hrs.

11.2.2 Supervision​‌

11.2.3 Juries​‌

11.2.4‌​‌ Educational and pedagogical outreach​​

• Juliette Janès gave a​​​‌ talk at (Speed meeting‌ between students and scientists),‌​‌ “Filles et Maths -​​ Rendez-vous des jeunes mathématiciennes​​​‌ et informaticiennes (RJMI)”. Inria‌ Paris, 23 Oct 2025.‌​‌
• Chloé Clavel gave a​​ talk at Lycée Louis​​​‌ Le Grand (Présentation orale‌ lycée, lors de la‌​‌ remise des prix Olympiades​​ de mathématiques), “La reconnaissance​​​‌ automatique des émotions dans‌ la voix”. , 25‌​‌ Jun 2025.
• Élodie Étienne​​ gave a talk at​​​‌ Lycée Paul Bert (Paris)‌ (Inria "Chiche !” programme),‌​‌ “Les maths comme passeport​​ pour les technologies du​​​‌ futur”. Lycée Paul Bert‌ (Paris), 4 Dec 2025.‌​‌
• Justine Cassell gave a​​ talk at Ecole Jeannine​​​‌ Manuel (IAckathon - Edition‌ collège), “IAckathon - Edition‌​‌ collège”. Paris, France, 19​​ Jun 2025.
• Djamé Seddah​​​‌ :
  • gave a talk‌ at Assocation Ro-BOTs (‌​‌ Ecole Bachelet, 14 rue​​ Alexandre Bachelet, Saint Ouen)​​​‌ (Intervention collègien 3e), Saint‌ Ouen, 21 Oct 2025.‌​‌
  • gave a talk at​​ locaux de l'association RoBoTic,​​​‌ école Bachelet (Saint Ouen)‌ (Témoignage métier - assoc.‌​‌ RoBoTIC), “Mon parcours et​​ l'intelligence artificielle”. Ecole Bachelet​​​‌ (Saint Ouen), 21 Oct‌ 2025.
• Gabrielle Le Bellier‌​‌ :
  • gave a talk​​ at (Speed meeting between​​​‌ students and scientists), “Filles‌ et Maths - Rendez-vous‌​‌ des jeunes mathématiciennes et​​ informaticiennes (RJMI)”. Inria Paris,​​​‌ 23 Oct 2025.
  • gave‌ a talk at (Inria‌​‌ “Chiche !" programme -​​ Chiche dans les mur”),​​​‌ “Mon parcours et les‌ biais dans les modèles‌​‌ de langue”. Inria Paris​​ , 11 Apr 2025.​​​‌
• Lydia Nishimwe :
  • gave‌ five talks as part‌​‌ of the Inria "Chiche​​ !” programme “La traduction​​​‌ automatique + Mon parcours”,‌ Lycée Maurice Ravel, Lycées‌​‌ Jacques Prévert et Marguerite​​ Yourcenar, Lycée de l'Essouriau,​​​‌ Lycée privé Charles de‌ Foucauld, Lycée Molière
  • animated‌​‌ an event Journée Citoyenne​​ IA (Interactive Demonstration), “L'apprenti​​​‌ illustrateur”. Hôtel de Ville‌ de Paris, 1 Jan‌​‌ 2025-1 Feb 2025.

11.3.1 Productions (articles,​​​‌ videos, podcasts, serious games,‌ ...)

11.3.2 Participation in​‌ Live events

International journals​​​‌

• 15 articleL.Lucie​ Chenain, A.Audrey​‌ Fabre, H.Hadrien​​ Titeux, G.Graça​​​‌ Morgado, K.Katia​ Youssov, C.Chloé​‌ Clavel and A.-C.Anne-Catherine​​ Bachoud-Lévi. Emotional speech​​​‌ markers of psychiatric disturbance​ in Huntington’s disease.​‌Frontiers in Psychiatry16​​August 2025, 1633492-1:1633492-20​​​‌HALDOIback to​ text
• 16 articleS.​‌Simon Gabay, A.​​Ariane Pinche, P.​​​‌Peter Nahon, A.​Alix Chagué, P.​‌Pauline Jacsont, É.​​Élodie Paupe, J.-C.​​​‌Jean-Claude Rebetez, M.​Maxime Humeau, C.​‌Christine Payot, T.​​Thibault Maillard, Y.​​​‌Yvan Jauregui, E.​Elina Leblanc and L.​‌Loraine Chappuis. Reading​​ Before It Can Be​​​‌ Read: Philological Reflections on​ Automatic Text Recognition for​‌ Modern French Manuscripts.​​Humanités numériques12December​​​‌ 2025HALDOIback​ to text
• 17 article​‌Y.Yanzhu Guo,​​ G.Guokan Shang and​​​‌ C.Chloé Clavel.​ Benchmarking Linguistic Diversity of​‌ Large Language Models.​​Transactions of the Association​​​‌ for Computational Linguistics13​November 2025, 1507-1526​‌HALDOIback to​​ text
• 18 articleE.​​​‌Emmanuelle Kempf, A.​A.T. Vu, E.​‌Eric Villemonte de La​​ Clergerie and R.Rémi​​​‌ Flicoteaux. 415P How​ to generate open source​‌ annotated cancer clinical datasets​​ with LLMs to support​​​‌ the development of smaller​ language models.ESMO​‌ Real World Data and​​ Digital Oncology10November​​​‌ 2025, 100611HAL​DOI
• 19 articleT.​‌ A.Tu Anh Nguyen​​, B.Benjamin Muller​​​‌, B.Bokai Yu​, M. R.Marta​‌ R. Costa-Jussa, M.​​Maha Elbayad, S.​​​‌Sravya Popuri, C.​Christophe Ropers, P.-A.​‌Paul-Ambroise Duquenne, R.​​Robin Algayres, R.​​​‌Ruslan Mavlyutov, I.​Itai Gat, M.​‌Mary Williamson, G.​​Gabriel Synnaeve, J.​​​‌Juan Pino, B.​Benoît Sagot and E.​‌Emmanuel Dupoux. SpiRit-LM:​​ Interleaved Spoken and Written​​​‌ Language Model.Transactions​ of the Association for​‌ Computational Linguistics13January​​ 2025, 30-52HAL​​​‌back to text
• 20​ articleA.Ariane Pinche​‌ and T.Thibault Clérice​​. Reconnaissance automatique d’écriture​​​‌ et sources historiques: Limites​ et nouvelles perspectives.​‌Studi francesi2062025​​, 357-366HALback​​ to text

Invited conferences​​​‌

• 21 inproceedingsA.Alix‌ Chagué. HTR-United schema‌​‌ for dataset description.​​2025 Workshop SCOOP -​​​‌ Source Codes of the‌ PastPrinceton , NJ,‌​‌ United StatesJune 2025​​HALback to text​​​‌

International peer-reviewed conferences

• 22‌ inproceedingsJ. O.Jesujoba‌​‌ O Alabi, I.​​ A.Israel Abebe Azime​​​‌, M.Miaoran Zhang‌, C.Cristina España-Bonet‌​‌, R.Rachel Bawden​​, D.Dawei Zhu​​​‌, D. I.David‌ Ifeoluwa Adelani, C.‌​‌ O.Clement Oyeleke Odoje​​, I.Idris Akinade​​​‌, I.Iffat Maab‌, D.Davis David‌​‌, S. H.Shamsuddeen​​ Hassan Muhammad, N.​​​‌Neo Putini, D.‌David Ademuyiwa, A.‌​‌Andrew Caines and D.​​Dietrich Klakow. AFRIDOC-MT:​​​‌ Document-level MT Corpus for‌ African Languages.Proceedings‌​‌ of the 2025 Conference​​ on Empirical Methods in​​​‌ Natural Language ProcessingProceedings‌ of the 2025 Conference‌​‌ on Empirical Methods in​​ Natural Language ProcessingSuzhou,​​​‌ ChinaAssociation for Computational‌ LinguisticsJanuary 2025,‌​‌ 27758-27794HALDOIback​​ to text
• 23 inproceedings​​​‌J.Javier Alejandro Lopetegui‌ Gonzalez, A.Arij‌​‌ Riabi and D.Djamé​​ Seddah. Common Ground,​​​‌ Diverse Roots: The Difficulty‌ of Classifying Common Examples‌​‌ in Spanish Varieties.​​VarDial 2025 - Twelfth​​​‌ Workshop on NLP for‌ Similar Languages, Varieties and‌​‌ Dialects co-located with COLING​​ 2025Abu Dhabi, United​​​‌ Arab EmiratesJanuary 2025‌HALback to text‌​‌
• 24 inproceedingsW.Wissam​​ Antoun, B.Benoît​​​‌ Sagot and D.Djamé‌ Seddah. ModernBERT or‌​‌ DeBERTaV3? Examining Architecture and​​ Data Influence on Transformer​​​‌ Encoder Models Performance.‌IJCNLP-AACL 2025 - 14th‌​‌ International Joint Conference on​​ Natural Language Processing (IJCNLP)​​​‌ and The 4th Conference‌ of the Asia-Pacific Chapter‌​‌ of the Association for​​ Computational Linguistics (AACL)Mumbai,​​​‌ IndiaNovember 2025HAL‌back to text
• 25‌​‌ inproceedingsA.Alisa Barkar​​, M.Mathieu Chollet​​​‌, M.Matthieu Labeau‌, B.Beatrice Biancardi‌​‌ and C.Chloé Clavel​​. Decoding Persuasiveness in​​​‌ Eloquence Competitions: An Investigation‌ into the LLM’s Ability‌​‌ to Assess Public Speaking​​.ICAART 2025 -​​​‌ 17th International Conference on‌ Agents and Artificial Intelligence‌​‌Proceedings of the 17th​​ International Conference on Agents​​​‌ and Artificial Intelligence -‌ Volume 3: ICAART, 538-546,‌​‌ 2025Porto, PortugalSCITEPRESS​​ - Science and Technology​​​‌ Publications2025, 538-546‌HALDOIback to‌​‌ text
• 26 inproceedingsR.​​Rachel Bawden, M.​​​‌Maud Bénard, E.‌Eric Villemonte de La‌​‌ Clergerie, J.José​​ Cornejo Cárcamo, N.​​​‌Nicolas Dahan, M.‌Manon Delorme, M.‌​‌Mathilde Huguin, N.​​Natalie Kübler, P.​​​‌Paul Lerner, A.‌Alexandra Mestivier, J.‌​‌Joachim Minder, J.-F.​​Jean-François Nominé, Z.​​​‌Ziqian Peng, L.‌Laurent Romary, P.‌​‌Panagiotis Tsolakis, L.​​Lichao Zhu and F.​​​‌François Yvon. MaTOS:‌ Machine Translation for Open‌​‌ Science.Proceedings of​​ Machine Translation Summit XX:​​​‌ Volume 220th Machine‌ Translation SummitVolume 2,‌​‌ Project Presentation PapersGeneva,​​ SwitzerlandJune 2025HAL​​​‌back to text
• 27‌ inproceedingsR.Rachel Bawden‌​‌ and B.Benoît Sagot​​​‌. RoCS-MT v2 at​ WMT 2025: Robust Challenge​‌ Set for Machine Translation​​.Proceedings of the​​​‌ Tenth Conference on Machine​ TranslationWMT 2025 -​‌ Tenth Conference on Machine​​ TranslationSuzhou, China2025​​​‌, 834–849HALback​ to textback to​‌ text
• 28 inproceedingsF.​​Florian Cafiero, L.​​​‌Lucence Ing, S.​Simon Gabay and T.​‌Thibault Clérice. "I​​ am too old for​​​‌ this style!" A stylometric​ benchmark of age effect​‌ on authorship attribution.​​CHR 2025 - 6th​​​‌ Conference on Computational Humanities​ ResearchLuxembourg, Luxembourg2025​‌HALback to text​​
• 29 inproceedingsT.Thibault​​​‌ Clérice and A.Ariane​ Pinche. Wauchier, Is​‌ That You? A multi-manuscript​​ authorship analysis of Saint​​​‌ Lambert's life.Anthology​ of Computers and the​‌ HumanitiesCHR 2025 -​​ 6th Conference on Computational​​​‌ Humanities Research3Luxembourg,​ Luxembourg2025, 149-165​‌HALDOIback to​​ text
• 30 inproceedingsF.​​​‌Fatima Dekmak, C.​Christian Khairallah and W.​‌Wissam Antoun. TutorMind​​ at BEA 2025 Shared​​​‌ Task: Leveraging Fine-Tuned LLMs​ and Data Augmentation for​‌ Mistake Identification.BEA​​ 2025 - 20th Workshop​​​‌ on Innovative Use of​ NLP for Building Educational​‌ ApplicationsProceedings of the​​ 20th Workshop on Innovative​​​‌ Use of NLP for​ Building Educational Applications (BEA​‌ 2025)Vienna, AustriaAssociation​​ for Computational LinguisticsJuly​​​‌ 2025, 1203-1211HAL​DOI
• 31 inproceedingsR.​‌Rasul Dent, P.​​Pedro Ortiz Suarez,​​​‌ T.Thibault Clérice and​ B.Benoît Sagot.​‌ Identifying Rare Languages in​​ Common Crawl Data is​​​‌ a Needles-in-a-Haystack Problem.​Findings of the Association​‌ for Computational Linguistics: EMNLP​​ 2025Suzhou, ChinaAssociation​​​‌ for Computational LinguisticsNovember​ 2025, 1460-1473HAL​‌DOIback to text​​back to text
• 32​​​‌ inproceedingsM.Matthieu Futeral​, C.Cordelia Schmid​‌, B.Benoît Sagot​​ and R.Rachel Bawden​​​‌. Towards Zero-Shot Multimodal​ Machine Translation.Findings​‌ of the Association for​​ Computational Linguistics: NAACL 2025​​​‌Findings of the Association​ for Computational Linguistics: NAACL​‌ 2025Albuquerque, New Mexico,​​ United States2025,​​​‌ 761–778HALback to​ textback to text​‌
• 33 inproceedingsM.Matthieu​​ Futeral, A.Armel​​​‌ Zebaze, P. O.​Pedro Ortiz Suarez,​‌ J.Julien Abadji,​​ R.Rémi Lacroix,​​​‌ C.Cordelia Schmid,​ R.Rachel Bawden and​‌ B.Benoît Sagot.​​ mOSCAR: A Large-scale Multilingual​​​‌ and Multimodal Document-level Corpus​.ACL 2025 -​‌ Findings of the Association​​ for Computational LinguisticsVienna,​​​‌ AustriaJuly 2025,​ 3461–3494HALDOIback​‌ to textback to​​ text
• 34 inproceedingsA.​​​‌Aina Garí Soler,​ M.Matthieu Labeau and​‌ C.Chloé Clavel.​​ Potentially Problematic Word Usages​​​‌ and How to Detect​ Them: A Survey.​‌14th Joint Conference on​​ Lexical and Computational Semantics​​​‌ (*SEM 2025)Suzhou, China​November 2025HALback​‌ to text
• 35 inproceedings​​A.Aina Garí Soler​​​‌, M.Matthieu Labeau​ and C.Chloé Clavel​‌. Toward the Automatic​​ Detection of Word Meaning​​​‌ Negotiation Indicators in Conversation​.EMNLP 2025 -​‌ 2025 Conference on Empirical​​ Methods in Natural Language​​ ProcessingSuzhou, ChinaNovember​​​‌ 2025HALback to‌ text
• 36 inproceedingsZ.‌​‌ M.Z. Melce Hüsünbeyi​​, D.Djamé Seddah​​​‌ and T.Tatjana Scheffler‌. Integrating Semantic Representations‌​‌ in a Cross-Modal Approach​​ to Fact-Checking.Proceedings​​​‌ of the 4th ACM‌ International Workshop on Multimedia‌​‌ AI against Disinformation, MAD'​​ 25MAD'25: 4th ACM​​​‌ International Workshop on Multimedia‌ AI against DisinformationChicago‌​‌ USA, United StatesACM​​June 2025, 17-27​​​‌HALDOIback to‌ text
• 37 inproceedingsJ.‌​‌Juliette Janes, R.​​Rachel Bawden, T.​​​‌Thibault Clérice, R.‌Rasul Dent, L.‌​‌Lucence Ing, O.​​Oriane Nédey and B.​​​‌Benoît Sagot. Encoding‌ language diversity in TEI:‌​‌ a description of regional​​ and non-standard languages in​​​‌ multilingual diachronic corpora.‌TEI Conference 2025Cracovie,‌​‌ PolandSeptember 2025HAL​​back to text
• 38​​​‌ inproceedingsB.Benjamin Kiessling‌. Version 5 of‌​‌ the Kraken ATR Engine​​ for the Humanities.​​​‌ICDAR2025 - 19th International‌ Conference on Document Analysis‌​‌ and RecognitionWuhan, China​​September 2025HALback​​​‌ to text
• 39 inproceedings‌C.Christine de Kock‌​‌, A.Arij Riabi​​, Z.Zeerak Talat​​​‌, M. S.Michael‌ Sejr Schlichtkrull, P.‌​‌Pranava Madhyastha and E.​​Ed Hovy. IYKYK:​​​‌ Using language models to‌ decode extremist cryptolects.‌​‌EACL 2026 - 19th​​ Conference of the European​​​‌ Chapter of the Association‌ for Computational LinguisticsRabat,‌​‌ MoroccoJune 2025HAL​​
• 40 inproceedingsT.Tom​​​‌ Kocmi, E.Ekaterina‌ Artemova, E.Eleftherios‌​‌ Avramidis, R.Rachel​​ Bawden, O.Ondřej​​​‌ Bojar, K.Konstantin‌ Dranch, A.Anton‌​‌ Dvorkovich, S.Sergey​​ Dukanov, M.Mark​​​‌ Fishel, M.Markus‌ Freitag, T.Thamme‌​‌ Gowda, R.Roman​​ Grundkiewicz, B.Barry​​​‌ Haddow, M.Marzena‌ Karpinska, P.Philipp‌​‌ Koehn, H.Howard​​ Lakougna, J.Jessica​​​‌ Lundin, C.Christof‌ Monz, K.Kenton‌​‌ Murray, M.Masaaki​​ Nagata, S.Stefano​​​‌ Perrella, L.Lorenzo‌ Proietti, M.Martin‌​‌ Popel, M.Maja​​ Popović, P.Parker​​​‌ Riley, M.Mariya‌ Shmatova, S.Steinthór‌​‌ Steingrímsson, L.Lisa​​ Yankovskaya and V.Vilém​​​‌ Zouhar. Findings of‌ the WMT25 General Machine‌​‌ Translation Shared Task: Time​​ to Stop Evaluating on​​​‌ Easy Test Sets.‌Proceedings of the Tenth‌​‌ Conference on Machine Translation​​WMT25 - Tenth Conference​​​‌ on Machine TranslationSuzhou,‌ ChinaAssociation for Computational‌​‌ LinguisticsNovember 2025,​​ 355-413HALDOIback​​​‌ to text
• 41 inproceedings‌M.Marc Lafon,‌​‌ Y.Yannis Karmim,​​ J.Julio Silva-Rodríguez,​​​‌ P.Paul Couairon,‌ C.Clément Rambour,‌​‌ R.Raphaël Fournier-Sniehotta,​​ I. B.Ismail Ben​​​‌ Ayed, J.Jose‌ Dolz and N.Nicolas‌​‌ Thome. ViLU: Learning​​ vision-language uncertainties for failure​​​‌ prediction.ICCV 2025.‌ International Conference on Computer‌​‌ VisionHonolulu, United States​​October 2025HAL
• 42​​​‌ inproceedingsM.Malik Marmonier‌, R.Rachel Bawden‌​‌ and B.Benoît Sagot​​. Explicit Learning and​​​‌ the LLM in Machine‌ Translation.Proceedings of‌​‌ the 2025 Conference on​​​‌ Empirical Methods in Natural​ Language ProcessingSuzhou, China​‌Association for Computational Linguistics​​March 2025, 31360-31410​​​‌HALDOIback to​ text
• 43 inproceedingsM.​‌Malik Marmonier, B.​​Benoît Sagot and R.​​​‌Rachel Bawden. A​ French Version of the​‌ OLDI Seed Corpus.​​Proceedings of the Tenth​​​‌ Conference on Machine Translation​ (WMT), Volume 2: Shared​‌ Task Papers,WMT 2025​​ - Tenth Conference on​​​‌ Machine TranslationSuzhou, China​Association for Computational Linguistics​‌2025, 1048-1060HAL​​DOIback to text​​​‌back to text
• 44​ inproceedingsA.Anh Ngo​‌, N.Nicolas Rollet​​, C.Catherine Pelachaud​​​‌ and C.Chloé Clavel​. "Mm, Wat?" Detecting​‌ Other-initiated Repair Requests in​​ Dialogue.EMNLP 2025​​​‌ - Conference on Empirical​ Methods in Natural Language​‌ ProcessingSuzhou, ChinaNovember​​ 2025HALback to​​​‌ text
• 45 inproceedingsC.​Célia Nouri, J.-P.​‌Jean-Philippe Cointet and C.​​Chloé Clavel. Graphically​​​‌ Speaking: Unmasking Abuse in​ Social Media with Conversation​‌ Insights.63rd Annual​​ Meeting of the Association​​​‌ for Computational Linguistics (ACL​ 2025)Vienna, AustriaJuly​‌ 2025HALback to​​ text
• 46 inproceedingsZ.​​​‌Ziqian Peng, R.​Rachel Bawden and F.​‌François Yvon. Self-Retrieval​​ from Distant Contexts for​​​‌ Document-Level Machine Translation.​Proceedings of the Tenth​‌ Conference on Machine Translation​​WMT 2025 - Conference​​​‌ on Machine TranslationSuzhou,​ ChinaNovember 2025HAL​‌
• 47 inproceedingsA.Arij​​ Riabi, V.Virginie​​​‌ Mouilleron, M.Menel​ Mahamdi, W.Wissam​‌ Antoun and D.Djamé​​ Seddah. Beyond Dataset​​​‌ Creation: Critical View of​ Annotation Variation and Bias​‌ Probing of a Dataset​​ for Online Radical Content​​​‌ Detection.COLING 2025​ - 31st International Conference​‌ on Computational LinguisticsAbu​​ Dhabi, United Arab Emirates​​​‌January 2025HALback​ to text
• 48 inproceedings​‌N. R.Nathaniel R​​ Robinson, C.Claire​​​‌ Bizon Monroc, R.​Rasul Dent, S.​‌Stefan Watson, K.​​Kenton Murray, R.​​​‌Raj Dabre, A.​Andre Coy and H.​‌Heather Lent. Findings​​ of the First Shared​​​‌ Task for Creole Language​ Machine Translation at WMT25​‌.WMT 2025 -​​ Tenth Conference on Machine​​​‌ TranslationSuzhou, ChinaAssociation​ for Computational LinguisticsNovember​‌ 2025, 520-531HAL​​DOIback to text​​​‌
• 49 inproceedingsS. C.​Sky Chenwei Wan,​‌ M.Matthieu Labeau and​​ C.Chloé Clavel.​​​‌ EmoDynamiX: Emotional Support Dialogue​ Strategy Prediction by Modelling​‌ MiXed Emotions and Discourse​​ Dynamics.Proceedings of​​​‌ the 2025 Conference of​ the Nations of the​‌ Americas Chapter of the​​ Association for Computational Linguistics:​​​‌ Human Language Technologies (Volume​ 1: Long Papers)Annual​‌ Conference of the Nations​​ of the Americas Chapter​​​‌ of the Association for​ Computational LinguisticsAlbuquerque, United​‌ StatesAssociation for Computational​​ LinguisticsApril 2025,​​​‌ 1678-1695HALDOI
• 50​ inproceedingsA. R.Armel​‌ Randy Zebaze, B.​​Benoît Sagot and R.​​​‌Rachel Bawden. Compositional​ Translation: A Novel LLM-based​‌ Approach for Low-resource Machine​​ Translation.Findings of​​​‌ the Association for Computational​ Linguistics: EMNLP 2025Findings​‌ of the Association for​​ Computational Linguistics: EMNLP 2025​​Suzhou, ChinaAssociation for​​​‌ Computational LinguisticsMarch 2025‌, 22328-22357HALDOI‌​‌back to text
• 51​​ inproceedingsA. R.Armel​​​‌ Randy Zebaze, B.‌Benoît Sagot and R.‌​‌Rachel Bawden. In-Context​​ Example Selection via Similarity​​​‌ Search Improves Low-Resource Machine‌ Translation.Findings of‌​‌ the Association for Computational​​ Linguistics: NAACL 2025Findings​​​‌ of the Association for‌ Computational Linguistics: NAACL 2025‌​‌Findings of the Association​​ for Computational Linguistics: NAACL​​​‌ 2025Albuquerque, United States‌Association for Computational Linguistics‌​‌2025, 1222-1252HAL​​DOIback to text​​​‌
• 52 inproceedingsA. R.‌Armel Randy Zebaze,‌​‌ B.Benoît Sagot and​​ R.Rachel Bawden.​​​‌ TopXGen: Topic-Diverse Parallel Data‌ Generation for Low-Resource Machine‌​‌ Translation.Findings of​​ the Association for Computational​​​‌ Linguistics: EMNLP 2025Findings‌ of the Association for‌​‌ Computational Linguistics: EMNLP 2025​​Suzhou, China2025,​​​‌ 22358–22381HALDOIback‌ to text