ChainCQG: Flow-Aware Conversational Question Generation

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 2061–2070 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 2061 ChainCQG: Flow-Aware Conversational Question Generation Jing Gu1,3 Mostafa Mirshekari1 Zhou Yu1,2 Aaron Sisto1 1 Searchable.ai 2 Columbia University 3 University of California, Davis {jing,mostafa,aaron}@searchable.ai zy2461@columbia.edu Abstract Conversational systems enable numerous valu- able applications, and question-answering is an important component underlying many of these. However, conversational question- answering remains challenging due to the lack of realistic, domain-specific training data. In- spired by this bottleneck, we focus on con- versational question generation as a means to generate synthetic conversations for training and evaluation purposes. We present a num- ber of novel strategies to improve conversa- tional flow and accommodate varying ques- tion types and overall fluidity. Specifically, we design ChainCQG as a two-stage archi- tecture that learns question-answer represen- tations across multiple dialogue turns using a flow propagation training strategy. ChainCQG significantly outperforms both answer-aware and answer-unaware SOTA baselines (e.g., up to 48% BLEU-1 improvement). Additionally, our model is able to generate different types of questions, with improved fluidity and corefer- ence alignment. 1 Introduction Conversational systems are important in many real- world applications, including personal assistants, educational tutors (Winkler et al., 2020), customer service (El Asri et al., 2017; Budzianowski et al., 2018), and increasingly, entertainment. A key com- ponent of these systems is the ability to interpret a search query and retrieve information from dif- ferent sources as naturally and efficiently as pos- sible. In analogous human interactions, such a search generally occurs through conversation. In this context, a conversation consists of a sequence of dialogue turns during which the search objec- tive becomes clearer over time. The applications mentioned above could benefit greatly from this type of multi-turn interaction, enabling conversa- tional agents to accurately predict intent, request additional information, and better understand am- biguous followup questions and comments. In an applied setting, meaningful and natural conversa- tions are important features of virtual entities as a means to establish trust and improve usability. Here, we are motivated by the challenging task of conversational question answering (CQA). Cur- rent open-source datasets such as CoQA (Reddy et al., 2019) and QuAC (Choi et al., 2018) pro- vide strong baselines for this task. However, these datasets have limited applicability in practical set- tings, because 1) they are created from domain- agnostic source material, and 2) they do not neces- sarily consider the full diversity of question types and vernacular that may be encountered in natu- ral dialogue. Creating realistic, domain-specific datasets to train CQA models is notoriously costly and time-consuming. As such, we focus on the re- lated task of question generation as a means to gen- erate synthetic conversational questions and subse- quently, create new datasets or augment existing ones. This will ultimately enable training CQA models in closed-loop, simulation environments, as well as allow machines to initiate dialogue and engage in information-seeking behavior. While QA models have been studied previ- ously (Zhu …

Content truncated to 500 words