How to Ask Good Questions? Try to Leverage Paraphrases

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6130–6140 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 6130 How to Ask Good Questions? Try to Leverage Paraphrases Xin Jia, Wenjie Zhou, Xu Sun, Yunfang Wu∗ MOE Key Lab of Computational Linguistics, School of EECS, Peking University {jemmryx, wjzhou013, xusun, wuyf}@pku.edu.cn Abstract Given a sentence and its relevant answer, how to ask good questions is a challenging task, which has many real applications. Inspired by human’s paraphrasing capability to ask questions of the same meaning but with di- verse expressions, we propose to incorporate paraphrase knowledge into question genera- tion(QG) to generate human-like questions. Specifically, we present a two-hand hybrid model leveraging a self-built paraphrase re- source, which is automatically conducted by a simple back-translation method. On the one hand, we conduct multi-task learning with sentence-level paraphrase generation (PG) as an auxiliary task to supplement paraphrase knowledge to the task-share encoder. On the other hand, we adopt a new loss function for diversity training to introduce more question patterns to QG. Extensive experimental results show that our proposed model obtains obvi- ous performance gain over several strong base- lines, and further human evaluation validates that our model can ask questions of high qual- ity by leveraging paraphrase knowledge. 1 Introduction Question generation (QG) is an essential task for NLP, which focuses on generating grammatical questions for given paragraphs or sentences. It plays a vital role in various realistic scenarios. For educational purposes, QG can create read- ing comprehension materials for language learn- ers (Heilman and Smith, 2010). For business use, QG can bring benefits to conversation sys- tems and chat-bots for effective communication with humans (Mostafazadeh et al., 2016). Be- sides, automatically-generated questions can be conversely used for constructing question answer- ing datasets to enhance reading comprehension sys- ∗ Corresponding author. Sentence: the next three drives of the game would end in punts. Answer: punts Reference question: what did the next three drives result in? Question generated by the baseline model: the next three drives of the game would end in what? Sentence: in ring theory, the notion of number is generally replaced with that of ideal. Answer: ring theory Reference question: in what theory is the idea of a number exchanged with that of an ideal? Question generated by the baseline model: in what theory is the notion of number replaced with that of ideal? Table 1: Real examples of generated questions from SQuAD. We highlight the paraphrase transitions be- tween sentences and questions. Human creates good questions by leveraging paraphrase knowledge, while the automatically generated questions just copy the original sentence, resulting in lower evaluation scores. tems (Tang et al., 2017; Duan et al., 2017; Xu et al., 2019; Zhang and Bansal, 2019). Recent neural network-based methods have achieved promising results on QG, most of which are based on the seq2seq attention framework (Du et al., 2017; Zhou et al., 2017; Gao et al., 2018; Kim et al., 2018; Zhou et al., …

Content truncated to 500 words