DIGING

Recent word embedding approches mainly focus on performance, sometimes at the cost of interpretability of time complexity. Meanwhile, interpretability is a major requirement to disseminate these technologies, in particular in the field of legal or medical tech. Furthermore, there is an urge to think about greener AI, considering the trade-off between performance and computationakl cost. With DIGING, we introduce a new graph based approach to learn word embeddings with a low-computational cost (linear-time complexity) while being performant. We also aim to embed words in a space made of physical latent structures uncovered in the graphs, namely communities. Interpretability of such embeddings would allow to consider their usage to deal with sensitive matters. We thus propose to experiment our approaches on two corpora from ANR projects the LIUM is a partner : ANTRACT and GEM. The first one is dedicated to the audiovisual history of France, and the second one to gender bias in medias. These corpus are temporel and we thus plan to consider incremental word embedding models based on our approach. To reach these goals, we ask for a post-doctoral stipend, a PhD student stipend, and support fundings.