Responding to customer queries automatically by customer reviews’ based Question Answering

Abstract

The entire world has been undergoing its own digital transformation over the past few decades as technology has advanced in leaps and bounds. Following this, an increase in the number of people using digital platforms for buying products online likewise increases the number of questions or enquiries posted about a product on an online shopping platform like Amazon on a day to day basis. Though we have gone completely digital in posting these questions, the answering of these questions is still manual. The forums are rarely active. By the time the user gets an answer to his question, either he has bought that product already through offline means or has lost interest in buying that product since it is time consuming. Moreover, the questions which are asked are mostly repetitive. At times the answers are already out there since they have already been given to some other user who had asked the same question. Also, lot of answers are embedded in the user reviews. Therefore, the answers can be extracted from the existing product reviews. This may lead to increase in sale and greater customer satisfaction as his query is resolved in much lower response time. We have review-based question answering systems that aim at answering the questions from the reviews given on the product by other customers. However, the existing systems have certain drawbacks due to the use of RNN, like missing attention mechanism etc. In this work, we enhance the performance of the existing review based QA systems by carrying out some prototypical experiments with the basic models of NLP and then moving towards more advanced Language Models while identifying and rectifying the shortcomings of the existing model. Further, in this work a thorough comparative analysis of the models and approaches that have been worked on is presented. We have enhanced the current state of the art existing review QA systems by using BERT, BART and also applied various heuristics for comparison. We achieved the best BLEU score of 0.58 by using BERT, which is an improvement of 0.19 on the current existing system.