ULMFiT for Question Classification in Question Answering Systems

Abstract

Question answering systems are used for retrieving relevant information that when a query is posted a direct response is provided rather than furnishing references containing answers. The main goal of these systems is to provide exact responses to users’ queries posed in natural language. The main aim of this paper is to perform question classification based on machine learning algorithms and transfer learning approach which helps in identifying the category to which a query belongs to. If the question asked by the user is “Will COVID end soon?”, then the Question answering system needs to retrieve information related to pandemic duration only. So, proper question categorization must be done by the system to help find similar questions that fall under the category of ” pandemic duration”. Then a comparison has been performed between two approaches. The experimental results on COVID-Q dataset demonstrate that the ULMFiT approach is better compared to TF-IDF and Word2Vec models combined with machine learning algorithms-Logistic Regression, MLP classifier, and SVM.