Answer-driven Deep Question Generation based on Reinforcement Learning

Proceedings of the 28th International Conference on Computational Linguistics, pages 5159–5170 Barcelona, Spain (Online), December 8-13, 2020 5159 Answer-driven Deep Question Generation based on Reinforcement Learning Liuyin Wang1 Zihan Xu1 Zibo Lin1 Hai-Tao Zheng1,∗ Ying Shen2,∗ 1Department of Computer Science and Technology, Tsinghua University 2School of Intelligent Systems Engineering, Sun Yat-Sen University {ly-wang19, xu-zh17,lzb18}@mails.tsinghua.edu.cn sheny76@mail.sysu.edu.cn, zheng.haitao@sz.tsinghua.edu.cn Abstract Deep question generation (DQG) aims to generate complex questions through reasoning over multiple documents. The task is challenging and underexplored. Existing methods mainly fo- cus on enhancing document representations, with little attention paid to the answer information, which may result in the generated question not matching the answer type and being answer- irrelevant. In this paper, we propose an Answer-driven Deep Question Generation (ADDQG) model based on the encoder-decoder framework. The model makes better use of the target an- swer as a guidance to facilitate question generation. First, we propose an answer-aware ini- tialization module with a gated connection layer which introduces both document and answer information to the decoder, thus helping to guide the choice of answer-focused question words. Then a semantic-rich fusion attention mechanism is designed to support the decoding process, which integrates the answer with the document representations to promote the proper handling of answer information during generation. Moreover, reinforcement learning is applied to inte- grate both syntactic and semantic metrics as the reward to enhance the training of the ADDQG. Extensive experiments on the HotpotQA dataset show that ADDQG outperforms state-of-the-art models in both automatic and human evaluations. 1 Introduction Neural question generation (QG) aims at generating specific answer related questions from a given doc- ument with a target answer based on deep neural networks. Its key applications include generating questions for reading comprehension (Du et al., 2017), enhancing question answering systems as a strat- egy of data augmentation (Tang et al., 2017; Zhang and Bansal, 2019) and helping digital assistants (e.g., Alexa, Cortana, Siri and Google Assistant) to start and continue a conversation. Various methods have been proposed for general QG (Zhou et al., 2017; Zhao et al., 2018; Kim et al., 2019; Zhang and Bansal, 2019; Tuan et al., 2020). However, most existing methods focus on generating questions relevant to only one fact without deep comprehension and reasoning. For example, Min et al. (2018) find that more than 80% of the questions in the widely adopted SQuAD dataset (Rajpurkar et al., 2016) are shallow and only relevant to information confined to a single sentence. Generating deep question which requires higher cognitive skills is rarely studied (Pan et al., 2020). These skills include a thorough understanding of the input sources and the ability to reason over disjoint and relevant contexts. This paper focuses on the task of deep question generation (DQG), which focuses on generating deep questions with multi-hop reasoning over document-level contexts. Previous work mainly focuses on the enhancement of document representations and obtains good performance. However, the answer information is also important since the generated questions should match the answer type and be answer- focused, and …

Content truncated to 500 words