PathQG: Neural Question Generation from Facts

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 9066–9075, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 9066 PathQG: Neural Question Generation from Facts Siyuan Wang1, Zhongyu Wei1,2∗, Zhihao Fan1, Zengfeng Huang1, Weijian Sun3, Qi Zhang4, Xuanjing Huang4 1School of Data Science, Fudan University, China 2Research Institute of Intelligent and Complex Systems, Fudan University, China 3Huawei Technologies Co., Ltd., China 4School of Computer Science, Fudan University, China {wangsy18,zywei,fanzh18,huangzf,qz,xjhuang}@fudan.edu.cn sunweijian@huawei.com Abstract Existing research for question generation en- codes the input text as a sequence of tokens without explicitly modeling fact information. These models tend to generate irrelevant and uninformative questions. In this paper, we ex- plore to incorporate facts in the text for ques- tion generation in a comprehensive way. We present a novel task of question generation given a query path in the knowledge graph con- structed from the input text. We divide the task into two steps, namely, query representa- tion learning and query-based question genera- tion. We formulate query representation learn- ing as a sequence labeling problem for identi- fying the involved facts to form a query and employ an RNN-based generator for question generation. We first train the two modules jointly in an end-to-end fashion, and further enforce the interaction between these two mod- ules in a variational framework. We construct the experimental datasets on top of SQuAD and results show that our model outperforms other state-of-the-art approaches, and the per- formance margin is larger when target ques- tions are complex. Human evaluation also proves that our model is able to generate rel- evant and informative questions.1 1 Introduction Question Generation (QG) from text aims to au- tomatically construct questions from textual in- put (Heilman and Smith, 2010). It receives increas- ing attentions from research communities recently, due to its broad applications in scenarios of dia- logue system and educational reading comprehen- sion (Piwek et al., 2007; Duan et al., 2017). It can also help to augment the question set to enhance the performance of question answering systems. ∗Corresponding author 1Our code is available at https://github.com/ WangsyGit/PathQG. (a) Machine generated questions (Qs) for an input text together with human generated ones (GTQs). Phrases underlined are the answers to ground-truth questions. (b) Knowledge graph constructed based on the input text shown in top sub-figure. Two colored ellipsoid are two query paths re- lated to two ground truth questions in sub-figure (a) respectively. Nodes in green are covered by ground-truth questions. Figure 1: A sample paragraph from SQuAD with machine generated questions (Zhou et al., 2017) (a), ground truth questions (a) and corresponding knowl- edge graph (b). Current QG systems mainly follow the sequence- to-sequence structure with an encoder for modeling the textual input and a decoder for text genera- tion (Du et al., 2017). These neural-based models have shown promising performance, however, they suffer from generating irrelevant and uninformative questions. Figure 1a presents two sample questions generated by a nueral QG model. Q2 contains ir- relevant information “Everton Fc”. Although Q1 is correct, …

Content truncated to 500 words