Complex Question Decomposition for Semantic Parsing

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 4477–4486 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 4477 Complex Question Decomposition for Semantic Parsing Haoyu Zhang†, Jingjing Cai‡, Jianjun Xu†, Ji Wang† †HPCL, College of Computer, National University of Defense Technology, China ‡Meituan-Dianping Group, China {zhanghaoyu10, jjxu, wj}@nudt.edu.cn caijingjing02@meituan.com Abstract In this work, we focus on complex question semantic parsing and propose a novel Hierar- chical Semantic Parsing (HSP) method, which utilizes the decompositionality of complex questions for semantic parsing. Our model is designed within a three-stage parsing archi- tecture based on the idea of decomposition- integration. In the first stage, we propose a question decomposer which decomposes a complex question into a sequence of sub- questions. In the second stage, we design an information extractor to derive the type and predicate information of these questions. In the last stage, we integrate the generated information from previous stages and gener- ate a logical form for the complex question. We conduct experiments on COMPLEXWE- BQUESTIONS which is a large scale com- plex question semantic parsing dataset, results show that our model achieves significant im- provement compared to state-of-the-art meth- ods. 1 Introduction Semantic parsing is a task which maps natural language utterances into logical forms such as SQL queries that can be executed based on re- lational databases or knowledge bases directly. Semantic parsing is a long-standing and difficult problem in natural language processing. In re- cent studies, researchers usually treat natural lan- guage descriptions/questions as input and use dif- ferent sequence-to-sequence frameworks to gener- ate logical forms (Xu et al., 2017; Dong and Lap- ata, 2016). However, these methods ignore the de- compositionality of a complex question which is usually composed of a set of sub-questions, the un- derstanding of each sub-question could contribute to the semantic parsing of the original complex question. Figure 1 gives an example of a complex ques- tion and its logical form. The related sub- questions in stage-1 and the corresponding pred- icate (relation) information of each sub-question in stage-2 could help to obtain the logical form of the complex question in stage-3. Stage 1 : Stage 2 : Stage 3 : When was Obama’s daughter born? Who is Obama’s daughter? When was #entity# born? people.person.children people.person.date_of_birth ��� ��� ��� ������� ������� �������� ������ � � ������� ������� ���������������� ��� Question: QD IE SP Figure 1: Example of question decomposition(QD), in- formation extraction(IE) and semantic parsing(SP). Question decomposition is important and many previous work utilize the decompositionality of complex questions to help question understand- ing. Kalyanpur et al. (2012) propose to use a suite of decomposition rules for question decomposi- tion. The drawback of rule-based methods is that it needs experts to design rules and the rules are usually with low coverage and hard to be extended to other domains and tasks. Talmor and Berant (2018) propose a neural question decomposition approach to answer complex questions. They use the pointer network (Vinyals et al., 2015) …

Content truncated to 500 words