Improving Dense FAQ Retrieval with Synthetic Training

Abstract

Frequently Asked Question (F AQ) retrieval is a valuable task which aims to find the most relevant question-answer pair from a FAQ dataset given a user query. Currently, most works implement F AQ retrieval considering the similarity between the query and the question as well as the relevance between the query and the answer. However, the query-answer relevance is difficult to model effectively due to the heterogeneity of query-answer pairs in terms of syntax and semantics. To alleviate this issue and improve retrieval performance, we propose a novel approach to consider answer information into F AQ retrieval by question generation, which provides high-quality synthetic positive training examples for dense retriever. Experiment results indicate that our method outperforms term-based BM25 and pretrained dense retriever significantly on two recently published COVID-19 F AQ datasets.