Models, Methods, and Multilingualism (M3)

The M3 team is interested in models and methods aimed at the fundamental description and automatic processing of natural language and its multilingual dimension. This involves the development of models and methods able to process, extract, and describe relevant characteristics of these systems from datasets. We consider languages in all their uses, with special focus on under-resourced languages and comparative approaches to language description (multilingual approaches). The team questions the interactions between models and systems: How does a model help our understanding of languages? How do sets of languages differ, and at what levels? How do we use generative models to produce controlled instances of given phenomena? The implementation of explainable and adapted models allows interaction between disciplines and promotes interdisciplinary work (computer science, language sciences, sociology, psychology). The research carried out in the M3 team falls into three main axes:

A- Models and Methods

This axis focuses on learning paradigms: developing data models and algorithms (models with many or few parameters, generative or not), with a view to their application to automatic language processing and languages as structured objects. These models are typically applied to the objects studied in the other two axes. Particular attention is paid to issues related to accessibility: reflecting on the specific methods to be implemented to develop inclusive technologies for our society. Effective, sober models, adapted to the representation of specific data, are particularly sought to promote explainable and responsible approaches to the data studied. These models make it possible to create hybrid solutions that try to control generative AIArtificial Intelligence. They also provide approaches that can be used to build more ethical systems.

B- Typology, Variation, and Universals in Languages

This axis focuses on describing the characteristics of linguistic systems based on corpora. This involves automatically applying typological schemes and language comparisons according to these characteristics. Considering variation is a major point, with work on diatopic, diastratic, diaphasic, or diachronic changes within languages, or linked to language contact of under-resourced or well-described languages. Syntactic, phonological, phonetic, articulatory, and prosodic systems are considered. The representation of the differences (in terms of distances or projected onto an atlas) between the systems studied is another highlight.

C- Contextualized Behaviors for Interaction

This axis models and describes performances during situated communicational interactions, at para- and extra-linguistic levels: whether for pragmatic functions (speech acts, attitudinal nuances), emotions (affective interaction, social emotions), nudges (gentle manipulation), vocal effort (Lombard speech, voice strength), etc. It aims to propose models of behavioral changes linked to these phenomena, in order to be able to detect them, measure their variation or dynamics, and categorize them. The analysis of the acoustic-linguistic parameters of the voice (parameters derived from models, glottal source, articulatory choices, etc.) makes it possible to link performances and functions.dèles, source glottique, choix articulatoires, etc.) permet de faire le lien entre les performances et les fonctions.

L’équipe se compose de 9 membres permanents (chercheurs CNRS, enseignants-chercheurs à l’Université Paris-Saclay), 6 doctorants, et 9 personnes ingénieures ou chercheurs CDD. Nous entretenons des liens avec les industriels (thèses en contrat CIFRE, projets de recherche) et organisons régulièrement des manifestations scientifiques (conférence TALN, ateliers et workshops scientifiques, etc.).