Word embeddings : models, time-based generalization, gender bias, interpretability
After an introduction to word embeddings models, we shall discuss the challenges related to their use in practice. First, we shall describe diachronic time-based word embeddings to monitor language evolution and neologisms. Second, because word embeddings are based on large textual corpora written by humans, these vectors reflect the human stereotypes. In particular, we consider gender bias. Third, we mention some existing approaches to bring interpretability to the word embeddings framework.
Word embeddings models, time-based generalization, gender bias, interpretability
Location: Inria Rennes – Bretagne Atlantique
Speaker: Nicolas Dugué