Models, Methods, and Multilingualism (M3)

The research focus of the Models, Methods, and Multilingualism team is on developing models and methods to help both the discovery of fundamental properties of language and the implementation of efficient systems to process it. We are interested in language in all its dimensions and modalities but strongly emphasize the multilingual dimension. The methods and models developed by the team are diverse by nature: computational (neural models, stochastic or symbolic methods), linguistic (language typology, linguistic diversity, and universals, affects), or societal (accessibility, nudges, language preservation, processing of underresourced languages and dialects). One of the team’s common aims is to relate language universals to the characteristics of language diversity and variation within a unified vision of linguistic and statistical (“automatic processing”) modeling of languages.

We can illustrate the team’s activity through a structured set of themes and sub-themes:

Universals in multilingual language modeling

Keywords: Linguistic diversity and universality in modeling; Representation of oral languages; Universal phonetic modeling and representation; Unified multilingual modeling and automatic identification of idiomaticity; Large multilingual and multimodal language models; Generative AI; Universal and cultural models of affects; Syntax of oral languages; Quantitative typology; Comparable corpora; Accessibility; Evaluation and resources; Multilingual generic systems: speech recognition, text generation, speech synthesis.

Methods and models for under-resourced languages

Keywords: Documentation of under-resourced languages; Scientific policies for endangered languages, Ethical and societal impact; Automatic processing of under-resourced languages; Massively multilingual models and interlingual transfer; Portability from a well-resourced to an under-resourced language.

Machine learning for NLP

Keywords: Machine learning and inference algorithms for structured prediction; Weak or unsupervised learning; Continuous learning; Representation learning and meta-learning; Learning in context of affective interactions.

Corpus linguistics, interlingual and intralingual variation

Keywords: Accents, dialects, and varieties: dialectometry (geoprosody) and linguistic cartography; Speaking styles; Variation of prosodic codes between languages and cultures (symbolic codes and socially coded attitudes); Expressive and multimodal prosody: illocutions, attitudes, social affects; Voice, voice strength, vocal quality, social uses.

Modeling of affective behaviors

Keywords: Automatic learning and detection of affective behaviors from paralinguistic and linguistic cues; Adaptation of large acoustic and linguistic models to emotion detection; Detection of abnormal behaviors and nudges in interaction; Ethical and societal impact of affects modeling and nudges.