Historically, natural language processing (NLP) has mainly focused on accessing the meaning of written or spoken contents. However, multiple other dimensions related to style, i.e. the way in which statements are realized, provide additional information to understand their underlying communication context (target audience, relationship between two interlocutors, emotional state or socio-cultural level of an author/speaker, modality...). Style processing is therefore an important issue to make artificial intelligence applications more "human", especially when these applications must generate linguistic contents for end-users.
This HDR summarizes my research activities over the last ten years, mostly within the Expression team, on the issues of variability and style in NLP. We consider the notion of style as the consistent use in texts or speech of various specific linguistic features (particular words, syntactic structures, pronunciations, etc.). Mainly, we seek to characterize these features, and then to introduce them into new sentences or utterances in order to mimic a desired style. These problems are declined on several instances of style (spontaneous speech, formality, language for children) as well as through a more generic conception of it. Beyond the different contributions, these two approaches lead us to question the respective roles of machine learning and human sciences towards a better processing of style in the future.
Please login in to get the URL (right panel) or contact Gwénolé Lecorvé.
Pierrette Bouillon, Université de Genève, Switerland, rapportrice
Mathew Magimai Doss, Idiap Research Institute, Switzerland, rapporteur
Iris Eshkol-Taravella, Université Paris-Nanterre, examinatrice
Élisa Fromont, Université de Rennes 1, examinatrice
Emmanuel Morin, Université de Nantes, examinateur