Performance Analysis of Machine Learning and Deep Learning Techniques for Answer Type extraction of Marathi Questions

Abstract

Question answering systems involve extraction of correct answers for natural language questions provided as input. Question classification is an important part of the question processing phase where natural language questions are categorized into predefined classes which specifies expected type of answer for the question. Extraction of answer type can be performed using machine learning and deep learning classification techniques which helps to reduce the list of possible correct answers for a question. In this paper we have compared various classification techniques which can be used for building a Marathi question classification system. Additionally, we have created a Marathi question classification dataset by translating existing TREC dataset available in English language. We observed that fine tuning the RoBERTa based monolingual language model for question classification was the best classification technique with accuracy of 91% in the coarse grained category of question classification and accuracy of 85% in the fine grained category of question classification.