A Transfer Learning Approach for Descriptive Question Answering System

Abstract

In this era of data explosion, where data is streaming at an exponential rate, efficient strategies of data processing have become an important aspect. The paper focuses on study of current "tate-of-the-art models" and proposes a Transfer learning approach with BERT model fine-tuned on SQuAD v2.0 dataset and improving the overall results based on speed and scaling factors and extending the reach of Question answering systems from generating one-line answers to generate descriptive answers. It is designed to handle user queries, whether it is web-based queries, reading comprehension tests, or Product related enquiries for satisfying user needs. The proposed model was experimented on a custom-made dataset to compare results with the original BERT model and a human-based evaluation method is proposed to evaluate the correctness of the descriptive answers.