Modeling What-to-ask and How-to-ask for Answer-unaware Conversational Question Generation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10785–10803 July 9-14, 2023 ©2023 Association for Computational Linguistics Modeling What-to-ask and How-to-ask for Answer-unaware Conversational Question Generation Xuan Long Do1,2,∗, Bowei Zou1, Shafiq Joty2,3†, Anh Tai Tran4, Liangming Pan5, Nancy F. Chen1, Ai Ti Aw1 1Institute for Infocomm Research (I2R), A*STAR, Singapore, 2Nanyang Technological University, Singapore, 3Salesforce AI, 4ByteDance, 5University of California, Santa Barbara {doxuanlong15052000, anhtai2672000}@gmail.com, liangmingpan@ucsb.edu, {zou_bowei, nfychen, aaiti}@i2r.a-star.edu.sg, srjoty@ntu.edu.sg Abstract Conversational Question Generation (CQG) is a critical task for machines to assist humans in fulfilling their information needs through con- versations. The task is generally cast into two different settings: answer-aware and answer- unaware. While the former facilitates the mod- els by exposing the expected answer, the latter is more realistic and receiving growing atten- tions recently. What-to-ask and how-to-ask are the two main challenges in the answer-unaware setting. To address the first challenge, existing methods mainly select sequential sentences in context as the rationales. We argue that the conversation generated using such naive heuris- tics may not be natural enough as in reality, the interlocutors often talk about the relevant contents that are not necessarily sequential in context. Additionally, previous methods decide the type of question (boolean/span-based) to be generated implicitly. Modeling the ques- tion type explicitly is crucial in this (answer- unaware) setting, as the answer which hints the models to generate a boolean or span-based question, is unavailable. To this end, we present SG-CQG, a two-stage CQG framework. For the what-to-ask stage, a sentence is selected as the rationale from a semantic graph that we con- struct, and extract the answer span from it. For the how-to-ask stage, a classifier determines the target answer type of the question via two ex- plicit control signals before generating and fil- tering. In addition, we propose Conv-Distinct, a novel evaluation metric for CQG, to evaluate the diversity of the generated conversation from a context. Compared with the existing answer- unaware CQG models, the proposed SG-CQG achieves state-of-the-art performance. 1 Introduction Building systems that can comprehend human speech and provide assistance to humans through ∗ Contribution during the internship at Institute for Info- comm Research. †Work done when the author was on leave from NTU. conversations is one of the main objectives in AI. Asking questions during a conversation is a cru- cial conversational behavior that helps AI agents communicate with humans more effectively (Allen et al., 2007; Li et al., 2016b). This line of research is known as Conversational Question Generation (CQG), which targets generating questions given the context and conversational history (Nakanishi et al., 2019; Pan et al., 2019a; Gu et al., 2021; Do et al., 2022). Compared to traditional single-turn question generation (Pan et al., 2019b), CQG is more challenging as the generated multi-turn ques- tions in a conversation need not only to be coherent but also follow a naturally conversational flow. Generally, there are two main settings for the CQG task: answer-aware and answer-unaware. In the answer-aware …

Content truncated to 500 words