Research projects

Coordinating Digital Humanities infrastructure activities in Europe (17 partners, 5 associated partners).

ATRIUM aims to empower Arts and Humanities scholars in their use of digital methods by facilitating access to a wide range of reusable workflows and interoperable, composable services offered by leading research infrastructures in the Arts and Humanities domain.

The MaTOS (Machine Translation for Open Science) project aims to develop new methods for the machine translation (MT) of complete scientific documents, as well as automatic metrics to evaluate the quality of these translations. Our main application target is the translation of scientific articles between French and English, where linguistic resources can be exploited to obtain more reliable translations, both for publication purposes and for gisting and text mining. However, efforts to improve MT of complete documents are hampered by the inability of existing automatic metrics to detect weaknesses in the systems and to identify the best ways to remedy them. The MaTOS project aims to address both of these issues.

SINNet proposes a paradigm shift for rendering conversational systems and social robotics a more acceptable and trustworthy technology even when using deep learning approaches. It will focus on the verbal component of the interaction, will target the agent-user social relationship, and model the behaviors indexing the state of the social relationship between agent and user, thus going beyond the analysis of the user’s positive and negative sentiments. It implies developing easy-to-adapt and easy-to-explain neural models able to analyse the user’s behavior contributing to user-agent co-construction processes such as the ones characterising the rapport with the agent, or the trust and affiliation in the agent, as well as to generate the agent’s answer fostering the user-agent social relationship. This SINNet project will establish interdisciplinarity as a core challenge by providing a shared formalism between complex (e.g., psychological or socio-linguistic) theories of social interactions and the underlying formalism in deep learning and language models.

The aim of TraLaLaM is to explore the use of large language models (LLMs) for machine translation, by asking two main questions: (i) in what scenarios can contextual information be effectively used via prompting? and (ii) for low-resource scenarios (with a focus on dialects and regional languages), can LLMs be effectively trained without any parallel data?

More than ever, with the increasing use of online video-conferencing solutions in daily professional interactions, public speaking skills are becoming crucial. The aim of this project is to obtain better insights into the best approaches allowing the practice of public speaking skills with technologically mediated tools. To this end, we will investigate different training environments (e.g. w/o a virtual/real audience) and different training approaches (e.g., modeling-based, feedback-based, simulation-based) to help users acquire, improve, and practice public speaking skills in full autonomy. For this purpose, different research challenges will be tackled to 1/ automatically learn, from different corpora, the multimodal cues correlated to the quality of public speaking; 2/ provide pedagogical activities rooted in coaching practice, taking a user-centered approach and 3/ provide a global evaluation of the training session as well as the specific behavioral characteristics to improve.

The text simplification algorithm developed within Cap'FALC is based on neural models for natural language processing. It will work similarly to a spell checker, which marks passages in a text, offers solutions but does not correct without a human validation step. The tool is intended to represent a valuable aid for disabled people responsible for transcribing texts in FALC, not to replace their intervention at all stages of the drafting; only their expertise can validate a text as being accessible and easy to read and understand.

ALMAnaCH is a member of the CORLI consortium on “corpora, languages and interactions” (B. Sagot is a member of the consortium’s board).

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.

The aim of the project is to make cancer data from health institutions accessible to all stakeholders involved for research and innovation purposes. The data at hand will be standardised and structured, in particular by extracting information from textual documents.

The Inria DEFI COLaF (Corpus and Tools for the Languages of France) aims to strengthen the ecosystem of automatic text and speech processing for the languages and speakers of France. To do this, it aims to create open datasets and use them to develop open-source models and tools.

The ambition of the ExcellencES TIERED project is to address the challenges of democratic systems in the face of environmental transformations and the digital transition, by producing outstanding scientific research, disseminating it within society, and training today's and tomorrow's decision-makers.

The project focuses on the production of transcriptions for literary manuscripts and public or private archives in Romance languages from the 11th to the 16th centuries. The main goal of the project is to produce training data and transcription models that are resilient to language and hand changes. HTRogenic is therefore envisaged as a building block for the infrastructure of Biblissima+ and the medieval philology of Romance languages: the project does not focus on a particular text or a small selection of texts, but on the contrary aims to produce examples of transcription capable of to constitute a representative sample. This sampling is based on specific criteria of language, script, genre and even dating.

SaLM is an interdisciplinary project between Inria Paris and Sciences Po that aims to redefine current NLP, LLM-based, algorithms by incorporating social contexts into their development and evaluation. It emphasises the importance of understanding language as a reflection of cultural and social identities. To explore this sociological dimension in NLP, the project will include two interrelated PhD projects about hate speech detection and cultural bias detection, gathering a mixed team of sociologists and NLPers to measure the role of the social dimension and prepare sociologically aware language models.

BACK IN TIME brings together the expertise of researchers in three fields - artificial intelligence, cryptography and history - to decipher encrypted historical letters, some of which have lain dormant for several centuries. Given the sheer number of pages and the variety of symbols and rules involved, our aim is to develop a software package to assist or even automate the deciphering of documents from ancient, medieval and modern history.

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.

CodeCommons is a two-year project building on the foundation of Software Heritage, the world’s largest public source code archive. Funded by the French government with academic partners in France and Italy, our mission is to expand and enhance the archive, consolidating critical, qualified information needed to create smaller, higher-quality datasets for the next generation of responsible AI tools. It prioritizes transparency and traceability, empowering model builders and users to respect creators’ rights while fostering a more sovereign and sustainable approach to AI development, massive software analysis, and reproducibility in research.

The PaRAMHTRS project advances large-scale experiments on medieval manuscripts (7th–15th centuries) in Latin and vernacular languages, leveraging HTR technology. It focuses on creating extensive corpora for culturomic studies and training models for ancient languages, while resolving abbreviations in HTR outputs. These efforts aim to enhance manuscript research and computational philology.

The Corpus Liberatum Linguae Graecae (CLLG) aims to advance open-access corpora for Ancient Greek and improve image-to-XML OCR methodologies. CLLG will develop a scalable and efficient digitisation pipeline designed to produce a reusable, high-quality corpus fully compatible with existing scholarly resources. The project places particular emphasis on the automatic structural annotation of Ancient Greek texts, adhering to the rigorous standards and practices of Ancient Greek philology.

The SCRIBE project is part of a strategic initiative for French digital sovereignty, aiming to develop large-scale language models (LLMs) tailored to the country's cultural, economic, and regulatory specificities. Led by a consortium of academic and industrial partners, it seeks to establish a robust, secure, and modular AI infrastructure, facilitating the industrialization and widespread adoption of LLMs. By integrating custom benchmarks, evaluation tools, and an open-source platform, SCRIBE aims to reduce reliance on foreign models while accelerating AI innovation and competitiveness in Europe.

PR[AI]RIE-PSAI (Paris School of AI), a follow-up to the PRAIRIE institute, is the largest of the AI Clusters established as part of the France 2030 national strategy. This ambitious project, led by PSL University, aims to create an internationally renowned school specialising in artificial intelligence. Its goal is to advance knowledge in AI, provide world-class higher education, and produce groundbreaking innovations in this field.

The Universal Dependencies project is an open community effort with over 300 contributors producing nearly 200 treebanks in over 100 languages. Universal Dependencies is a framework for consistent annotation of grammar (parts of speech, morphological features, and syntactic dependencies) across different human languages. With a release every 6 months since 2017, the UD framework is the de-facto standard for syntactic structures representations and of course the basis for most if not all supervised neural parsers.

Computational models of personality have largely ignored the subtleties of expressions of personality, and have assumed that individuals (and therefore embodied conversational agents) should have one fixed personality (such as introvert or extrovert), despite the fact that all of the psychological evidence demonstrates that the expression of one's personality changes as a function of who one is speaking to, what one is speaking about, the personalities of the other individuals in a conversation, one's culture, one's age, etc. A new model is needed that takes into account the malleability of personality and therefore the need for conversational agents to also have malleable representations of personality so as to be maximally effective in working with people.