Resource-light multilingual natural language interface

Abstract

The aim of this thesis is to explore the use of cross-lingual word embeddings (CLWEs) while building a light-resource and language-independent retrieval chatbot. For this purpose, by using MUSE CLWEs, we created a multilingual framework in which it’s possible to quickly implement several different topics (bots) that can reply to any desired argument. We tested the framework using four different topics (weather, travel, Wikipedia, time) and three different languages (Italian, Spanish, English). Subsequently, we evaluated both the NLU module of the bots we implemented upon Italian and also the general framework, presenting our work in a three-days workshop on the topic of “Chatbots” at the University of Tartu. The results of our work show that, through the combination of machine translation and CLWEs, we can achieve high accuracy in the responses returned by the conversational agent while the approach based only on multilingual CLWEs requires further research to be as performative.