S2SQL: Injecting Syntax to Question-Schema Interaction Graph Encoder for Text-to-SQL Parsers

Findings of the Association for Computational Linguistics: ACL 2022, pages 1254 - 1262 May 22-27, 2022 c⃝2022 Association for Computational Linguistics S2SQL: Injecting Syntax to Question-Schema Interaction Graph Encoder for Text-to-SQL Parsers Binyuan Hui 1, Ruiying Geng 1, Lihan Wang 1,2,3 ∗, Bowen Qin 1,2,3 ∗, Yanyang Li 4, Bowen Li 1, Jian Sun 1 and Yongbin Li 1 † 1 Alibaba Group, 2 University of Chinese Academy of Sciences 3 Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences 4 The Chinese University of Hong Kong {binyuan.hby,ruiying.gry,yanjin.lbw,jian.sun,shuide.lyb}@alibaba-inc.com Abstract The task of converting a natural language ques- tion into an executable SQL query, known as text-to-SQL, is an important branch of seman- tic parsing. The state-of-the-art graph-based encoder has been successfully used in this task but does not model the question syntax well. In this paper, we propose S2SQL, injecting Syntax to question-Schema graph encoder for Text-to-SQL parsers, which effectively lever- ages the syntactic dependency information of questions in text-to-SQL to improve the perfor- mance. We also employ the decoupling con- straint to induce diverse relational edge em- bedding, which further improves the network’s performance. Experiments on the Spider and robustness setting Spider-Syn demonstrate that the proposed approach outperforms all existing methods when pre-training models are used, resulting in a performance ranks first on the Spider leaderboard. 1 Introduction Relational databases are ubiquitous and store a great amount of structured information. The in- teraction with databases often requires expertise on writing structured code like SQL, which is not friendly for users who are not proficient in query languages. Text-to-SQL aims to automatically translate natural language questions into executable SQL statements (Zelle and Mooney, 1996; Zettle- moyer and Collins, 2005; Wong and Mooney, 2007; Zettlemoyer and Collins, 2007; Berant et al., 2013; Li and Jagadish, 2014; Yaghmazadeh et al., 2017; Iyer et al., 2017). Recently, a large-scale, multi-table, realistic text- to-SQL benchmark, Spider (Yu et al., 2018), has been released. The most effective and popular en- coder architecture on Spider is the question-schema interaction graph (Wang et al., 2020). Built on that, ∗ Work done during an internship at Alibaba DAMO. † Corresponding author. Show the date of the transcript which shows at least number of results, also list the id. Schema Pred SQL Question Gold SQL SELECT transcript_date, FROM Transcript_Contents AS T1 JOIN … SELECT transcript_date, transcript_id FROM Transcript_Contents AS T1 JOIN … transcript_date transcript_contents transcript_id Column Table Column dep obj Figure 1: A typical bad case of the current graph-based methods. If the structure of question (dependencies) are not considered, the wrong SQL will be generated even if the linking is correct. many state-of-the-art models have been further de- veloped (Chen et al., 2021; Cao et al., 2021). It jointly models natural language question and struc- tured database schema information, and uses some pre-defined relationships to carve out the interac- tion between them. However, we observed that the current graph-based model yet has two major limitations. Syntactic Modelling. Jointly modeling syntax and semantics is a core problem in NLP. …

Content truncated to 500 words