Dynamic and Interpretable Graph-based word embeddINGs (DIGING)
Date: 01/2022 - 12/2025
Call: ANR JCJC
Recent approaches to embedding learning have focused on results, often at the expense of interpretability and algorithmic complexity. Yet, interpretability is a necessary pre-requisite for the deployment of such technologies when they are used in sensitive domains such as law or medicine. Moreover, ecological imperatives create an urgent need to think about efficient and computationally efficient systems. With DIGING, we propose a new powerful and computationally efficient approach for the construction of interpretable lexical plots based on the theory of complex networks. With this original approach, the goal is to build vectors integrating polysemy natively by plunging words into a space with interpretable dimensions.
The interpretability of such embeddings allows us to consider applications related to sensitive domains and society topics. For instance, we propose to apply the automatic methods developed in the project on corpora from two ANR projects in which LIUM is a partner: ANTRACT and GEM, the former related to the evolution of France through its audiovisual history, and the latter studying gender representations in the media. The interpretability of the models developed with DIGING is particularly suitable for the mining of such corpora, especially in a temporal context, considering the detection and characterization of semantic neologisms.