Effectiveness of Data Augmentation to Identify Relevant Reviews for Product Question Answering

With the rapid growth of e-commerce and an increasing number of questions posted on the Question Answer (QA) platforms of e-commerce websites, there is a need for providing automated answers to questions. In this paper, we use transformer-based review ranking models which provide a ranked list of reviews as a potential answer to a new question. Since no explicit training data is available, we exploit the product reviews along with available QA pairs to learn a relevance function between a question and a review sentence. Further, we present a data augmentation technique by fine-tuning the T5 model to generate new questions from customer reviews by considering the summary of the review as an answer and the review as the document. We conduct experiments on a real-world dataset from three categories in Amazon.com. To assess the performance of the models, we use the annotated question review dataset from RIKER [13]. Experimental results show that Deberta-RR model with the augmentation technique outperforms the current state-of-the-art model by 5.84%, 4.38%, 3.96%, and 2.96% on average in nDCG@1, nDCG@3, nDCG@5, and nDCG@10, respectively.