Semantic Malayalam Dialogue System For Covid-19 Question Answering Using Word Embedding And Cosine Similarity

Abstract

Covid-19 is a global pandemic, has affected millions of people physically and mentally. The dynamic and rapidly growing situation with COVID-19 made it more difficult to discourse accurate and authoritative information about the disease, in most of the Indian local languages like Malayalam. To resolve this issue, here we propose a semantic Malayalam Dialogue System for COVID-19 related Question Answering. This is a user-friendly knowledge system to automatically deliver relevant answers to COVID-19 related queries in the Malayalam language. The proposed system proceeds in three stages; Document pre-processing, Semantic modelling using word embedding and Answer Retrieval. The NLP techniques are used for document processing, word embedding - CBOW and Skip Gram methods, Neural Network models are used for Semantic Modelling and finally, a cosine similarity measure is used to map and retrieve the answers for the user's queries. The experiment was conducted with our own Malayalam dataset; and compared the performance of two Word2Vec algorithms - CBOW and Skip Gram. The result, with our data set, shows that Skip-Gram is more efficient than CBOW and CBOW is faster than the Skip Gram model.