Recently, multilingual NLP models has gained attention in the NLP field. They are supposed to handle multiple languages in one single model, but one of the main problems is the huge diversity of human languages. We try to cope with this problem in the light of linguistic typology, which offers a systematic comparison of the world’s languages in terms of a variety of linguistic properties.
Download PosterCross entropy loss only evaluates sentences on the token level and is unable to handle synonyms or changes in sentence structure. For this reason, we propose to evaluate output sentences with more flexible criteria such as their Semantic Textual Similarity (STS) with ground truth sentences, then use Reinforcement Learning (RL) with estimated STS scores as reward.
Download Poster