Seminar from Martin Atangana Ongolo, PhD student at Université Yaoundé 1 and ESPERANTO secondee at LIUM

Date: 31/05/2024
Time: 14h00
Localization: IC2, Boardroom
Speaker: Martin Atangana Ongolo
 
 

How to use multilingual embeddings for LRL(Low resources Languages)

The aim of this seminar is to present the problem of mapping quality for Low resources languages. The representations obtained with plunging methods are now widely used for NLP tasks such as machine translation, lexicon induction, multilingual information retrieval, etc.

The idea is to improve these representations by exploiting those already identified as being of very high quality (plunking on richly endowed languages).

We first present experiments carried out for machine translation with a neural architecture without multilingual embedding. We then consider the multilingual embeddings of the language pairs (French – Ewondo) and (Yoruba – Ewondo). This second embedding produces better performance for the machine translation task than the first. We then raise the question of how to measure the quality of multilingual embeddings, and explore some initial approaches.