Learning to Identify Follow-Up Questions in Conversational Question Answering

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 959–968 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 959 Learning to Identify Follow-Up Questions in Conversational Question Answering Souvik Kundu, Qian Lin, and Hwee Tou Ng Department of Computer Science, National University of Singapore souvik@u.nus.edu, qlin@u.nus.edu, nght@comp.nus.edu.sg Abstract Despite recent progress in conversational ques- tion answering, most prior work does not fo- cus on follow-up questions. Practical conver- sational question answering systems often re- ceive follow-up questions in an ongoing con- versation, and it is crucial for a system to be able to determine whether a question is a follow-up question of the current conversa- tion, for more effective answer finding subse- quently. In this paper, we introduce a new follow-up question identification task. We pro- pose a three-way attentive pooling network that determines the suitability of a follow-up question by capturing pair-wise interactions between the associated passage, the conversa- tion history, and a candidate follow-up ques- tion. It enables the model to capture topic continuity and topic shift while scoring a par- ticular candidate follow-up question. Experi- ments show that our proposed three-way atten- tive pooling network outperforms all baseline systems by significant margins. 1 Introduction Conversational question answering (QA) mimics the process of natural human-to-human conversa- tion. Recently, conversational QA has gained much attention, where a system needs to answer a series of interrelated questions from an associated text passage or a structured knowledge graph (Choi et al., 2018; Reddy et al., 2019; Saha et al., 2018). However, most conversational QA tasks do not ex- plicitly focus on requiring a model to identify the follow-up questions. A practical conversational QA system must possess the ability to understand the conversation history well, and to identify whether the current question is a follow-up of that partic- ular conversation. Consider a user who is trying to have a conversation with a machine (e.g., Siri, Google Home, Alexa, Cortana, etc). First, the user asks a question and the machine answers it. When Passage: ...script for Verhoeven’s first American film, Flesh and Blood (1985), which starred Rutger Hauer and Jennifer Jason Leigh. Verhoeven moved to Hol- lywood for a wider range of opportunities in filmmak- ing. Working in the U.S. he made a serious change in style, directing big-budget, very violent, special-effects- heavy smashes RoboCop and Total Recall. RoboCop, for ... Verhoeven followed those successes with the equally intense and provocative Basic Instinct (1992) ... received two Academy Awards nominations, for Film Editing and for Original Music . . . Conversation history: Q: What was the first film Verhoeven did in the US? A: Flesh and Blood Q: What genre of films did he make? A: big-budget, very violent, special-effects-heavy smashes Candidate follow-up question examples: What year did his first film debut? – Valid Did he make any films during his final years? – Invalid What did she do after her debut film? – Invalid Figure 1: Examples illustrating the follow-up question identification task. the user asks …

Content truncated to 500 words