Mention Extraction and Linking for SQL Query Generation

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 6936–6942, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 6936 Mention Extraction and Linking for SQL Query Generation Jianqiang Ma∗, Zeyu Yan∗, Shuai Pang, Yang Zhang, Jianping Shen {majianqiang554,yanzeyu751,pangshuai550,zhangyang147,shenjianping324}@pingan.com.cn AI Team, Ping An Life Insurance Company of China, Ltd Abstract On the WikiSQL benchmark, state-of-the-art text-to-SQL systems typically take a slot- filling approach by building several dedicated models for each type of slots. Such mod- ularized systems are not only complex but also of limited capacity for capturing inter- dependencies among SQL clauses. To solve these problems, this paper proposes a novel extraction-linking approach, where a unified extractor recognizes all types of slot men- tions appearing in the question sentence be- fore a linker maps the recognized columns to the table schema to generate executable SQL queries. Trained with automatically generated annotations, the proposed method achieves the first place on the WikiSQL benchmark. 1 Introduction Text-to-SQL systems generate SQL queries accord- ing to given natural language questions. Text- to-SQL technology is very useful as it can em- power humans to naturally interact with relational databases, which serve as foundations for the dig- ital world today. As a subarea of semantic pars- ing (Berant et al., 2013), text-to-SQL is known to be difficult due to the flexibility in natural language. Recently, by the development of deep learning, significant advances have been made in text-to- SQL. On the WikiSQL (Zhong et al., 2018) bench- mark for multi-domain, single table text-to-SQL, state-of-the-art systems (Hwang et al., 2019; He et al., 2019) can predict more than 80% of en- tire SQL queries correctly. Most of such systems take a sketch-based approach (Xu et al., 2018) that builds several specialized modules, each of which is dedicated to predicting a particular type of slots, such as the column in SELECT, or the filter value in WHERE. Such dedicated modules are complex and often fall short of capturing inter-dependencies ∗Equal contributions. among SQL sub-clauses, as each type of slots is modeled separately. To deal with these drawbacks, this paper formulates text-to-SQL as mention ex- traction and linking problems in a sequence la- beling manner (Section 2). In this new formula- tion, the key to synthesizing SQL is to extract the mentions of SQL slots and the relations between them. Consider the question and its correspond- ing SQL query in example (1), with the headers in the schema being {LANE, NAME, NATIONALITY, SPLIT (50M), TIME}. (1) a. Question: What is the total sum of 50m splits for Josefin Lillhage in lanes above 8? b. SQL: SELECT SUM (Split (50m)) FROM some table WHERE Name = ‘Josefin Lillhage’ AND Lane > 8 We can see that many SQL elements, or slots, such as column names of SPLIT (50M) and LANE, values like “Josefin Lillhage” and 8, as well as operators > are mentioned with words similar in form and/or meaning. Moreover, the relations be- tween the slot mentions, such as <lanes, above, 8> forming a filter …

Content truncated to 500 words