From Paraphrasing to Semantic Parsing: Unsupervised Semantic Parsing via Synchronous Semantic Decoding

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 5110–5121 August 1–6, 2021. ©2021 Association for Computational Linguistics 5110 From Paraphrasing to Semantic Parsing: Unsupervised Semantic Parsing via Synchronous Semantic Decoding Shan Wu1,3, Bo Chen1, Chunlei Xin1,3, Xianpei Han1,2,∗, Le Sun1,2,∗, Weipeng Zhang4, Jiansong Chen4, Fan Yang4, Xunliang Cai4 1Chinese Information Processing Laboratory 2State Key Laboratory of Computer Science Institute of Software, Chinese Academy of Sciences, Beijing, China 3University of Chinese Academy of Sciences, Beijing, China 4Meituan {wushan2018,chenbo,xianpei,sunle}@iscas.ac.cn, xinchunlei20@mails.ucas.ac.cn, {zhangweipeng02,chenjiansong,yangfan79,caixunliang}@meituan.com Abstract Semantic parsing is challenging due to the structure gap and the semantic gap between utterances and logical forms. In this paper, we propose an unsupervised semantic pars- ing method – Synchronous Semantic Decod- ing (SSD), which can simultaneously resolve the semantic gap and the structure gap by jointly leveraging paraphrasing and grammar- constrained decoding. Specifically, we refor- mulate semantic parsing as a constrained para- phrasing problem: given an utterance, our model synchronously generates its canonical utterance1 and meaning representation. Dur- ing synchronous decoding: the utterance para- phrasing is constrained by the structure of the logical form, therefore the canonical utterance can be paraphrased controlledly; the seman- tic decoding is guided by the semantics of the canonical utterance, therefore its logical form can be generated unsupervisedly. Experimen- tal results show that SSD is a promising ap- proach and can achieve competitive unsuper- vised semantic parsing performance on multi- ple datasets. 1 Introduction Semantic parsing aims to translate natural lan- guage utterances to their formal meaning repre- sentations, such as lambda calculus (Zettlemoyer and Collins, 2005; Wong and Mooney, 2007), FunQL (Kate et al., 2005; Lu et al., 2008), and SQL queries. Currently, most neural semantic parsers (Dong and Lapata, 2016; Chen et al., 2018b; Zhao et al., 2020; Shao et al., 2020) model semantic parsing as a sequence to sequence trans- lation task via encoder-decoder framework. ∗Corresponding Author 1Canonical utterances are pseudo-language representa- tions of logical forms, which have the synchronous structure of logical forms.(Berant and Liang, 2014; Xiao et al., 2016; Su and Yan, 2017; Cao et al., 2020) Paraphrase Model Semantic Parsing Grammar Paraphrasing Synchronous Semantic Decoding (Our Method) Structure Constraints Semantic Consistency Logical Form: Answer ( count ( river ( traverse_2 ( state0 ) ) ) ) Canonical Utterance: What is the number of river traverse state0 Utterance: How many rivers run through state0 Semantic Consistency Structure Constraints Figure 1: Different from previous staged methods (in- dicated by gray lines), our method generates canonical ut- terance and logical form synchronously. The semantic gap and the structure gap are simultaneously resolved by jointly leveraging paraphrasing and grammar-constrained decoding. Thus, our synchronous decoding employs both the semantic and the structure constraints to solve unsupervised semantic parsing. Semantic parsing is a challenging task due to the structure gap and the semantic gap between natural language utterances and logical forms. For structure gap, because utterances are usu- ally word sequences and logical forms are usually trees/graphs constrained …

Content truncated to 500 words