Retrieval-Based Systems for Open-domain Question Answering

Abstract

Question answering is a popular natural language processing task in which a question answering (QA) system is responsible for automatically producing relevant and factually accurate natural language answers to questions formulated in natural language. Open-domain QA refers to the general setting in which the QA system is not provided the evidence required to answer the questions and in which the questions are not restricted to any particular domain. This is a challenging task as it requires open-domain QA systems to have access to a vast amount of factual knowledge pertaining to a diverse range of topics and real-world entities. We first motivate the need for retrieval in open-domain QA, identifying several distinct advantages it has over competing retrieval-free methods. We then survey the most prominent and influential approaches for retrieval-based open-domain QA, focusing on the popular retriever-reader framework. In particular, we describe and compare the various designs proposed for retriever and reader models and detail the methods to train and evaluate such open-domain QA systems. We finally discuss some interesting novel directions for future research in retrieval-based open-domain QA.