Seminare from Paulin Mélatagia, Université Yaoundé 1, ESPERANTO secondee at LIUM


Date: 13/06/2024
Time: 14h00
Localization: IC2, Boardroom
Speaker: Paulin Mélatagia

Extend self supervised representation models with features engineering for NLP of African Languages


Most of the African languages are low-resourced and their linguistic characteristics are different from those of the main languages used by machine learning models. Our work focuses on providing better representations of text and speech for these languages. I will present our approaches of features engineering (abstract syntax tree, distributional representation, …) and representation learning (multilingual, contrastive, …) to improve the quality of NLP models for African languages.

We propose to combine the two classes of representations to enhance the integration of syntactic and semantic properties of African languages in NLP applications. I will also introduce recents research questions on explainability using graph neural networks clustering-based models for the diarization.