FAQ-Based Question Answering Systems with Query-Question and Query-Answer Similarity

Abstract

A Frequently Asked Question (FAQ) Answering System maximizes knowledge access by enabling users to request a natural language question using the FAQ database. Retrieving FAQs is challenging due to the linguistic difference between a query and a question-answer pair. This work explores methods to improve on this linguistic gap in FAQ retrieval of the Question Answering System. The task is to retrieve frequently asked question-answer pairs (FAQ pairs) from the database that are related to the user’s query, thus providing answers to the user. We do so by leveraging natural language processing models like BERT and SBERT and ranking functions like BM25. The best results are obtained when BERT is trained in a triplet fashion (question, paraphrase, non-matching question) and combined with the BM25 model, which compares query with FAQ question answer concatenation.