Learn to Resolve Conversational Dependency: A Consistency Training Framework for Conversational Question Answering

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 6130–6141 August 1–6, 2021. ©2021 Association for Computational Linguistics 6130 Learn to Resolve Conversational Dependency: A Consistency Training Framework for Conversational Question Answering Gangwoo Kim Hyunjae Kim Jungsoo Park Jaewoo Kang† Korea University {gangwoo kim,hyunjae-kim}@korea.ac.kr {jungsoo park,kangj}@korea.ac.kr Abstract One of the main challenges in conversational question answering (CQA) is to resolve the conversational dependency, such as anaphora and ellipsis. However, existing approaches do not explicitly train QA models on how to resolve the dependency, and thus these mod- els are limited in understanding human dia- logues. In this paper, we propose a novel framework, EXCORD (Explicit guidance on how to resolve Conversational Dependency) to enhance the abilities of QA models in com- prehending conversational context. EXCORD first generates self-contained questions that can be understood without the conversation history, then trains a QA model with the pairs of original and self-contained questions using a consistency-based regularizer. In our exper- iments, we demonstrate that EXCORD signifi- cantly improves the QA models’ performance by up to 1.2 F1 on QuAC (Choi et al., 2018), and 5.2 F1 on CANARD (Elgohary et al., 2019), while addressing the limitations of the existing approaches.1 1 Introduction Conversational question answering (CQA) involves modeling the information-seeking process of hu- mans in a dialogue. Unlike single-turn question answering (QA) tasks (Rajpurkar et al., 2016; Kwiatkowski et al., 2019), CQA is a multi-turn QA task, where questions in a dialogue are context- dependent;2 hence they need to be understood with the conversation history (Choi et al., 2018; Reddy et al., 2019). As illustrated in Figure 1, to answer † Corresponding author 1Our models and code are available at: https://github.com/dmis-lab/excord 2While the term “context” usually refers to the evidence document from which the answer is extracted, in CQA, it refers to conversational context. Title : Leonardo da Vinci Conversation History with his pupils Salai and Melzi. a1 Who were his pupils? q Leonardo's most intimate relationships a2 Was he close to his pupils? q Was he close with anyone else? q Was Leonardo da Vinci close with anyone else other than his pupils Salai and Melzi? q Self-contained Q Question Rewrite 1 2 3 3 Figure 1: An example of the QuAC dataset (Choi et al., 2018). Owing to linguistic phenomena in human con- versations, such as anaphora and ellipsis, the current question q3 should be understood based on the conver- sation history: q1, a1, q2, and a2. Question q3 can be reformulated as a self-contained question ˜q3 via a ques- tion rewriting (QR) process. the current question “Was he close with anyone else?,” a model should resolve the conversational dependency, such as anaphora and ellipsis, based on the conversation history. A line of research in CQA proposes the end-to- end approach, where a single QA model jointly encodes the evidence document, the current ques- tion, and the whole conversation history (Huang et al., 2018; Yeh and Chen, …

Content truncated to 500 words