Service-oriented Text-to-SQL Parsing

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2218–2222 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 2218 Service-oriented Text-to-SQL Parsing Wangsu Hu, Jilei Tian BMW Technology Corporation / 540 W Madison St 2400, Chicago, IL 60661 wangsu.hu@bmwna.com, jilei.tian@bmwna.com Abstract The information retrieval from relational database requires professionals who has an un- derstanding of structural query language such as SQL. TEXT2SQL models apply natural lan- guage inference to enable user interacting the database via natural language utterance. Cur- rent TEXT2SQL models normally focus on generating complex SQL query in a precise and complete fashion while certain features of real-world application in the production envi- ronment is not fully addressed. This paper is aimed to develop a service-oriented Text-to- SQL parser that translates natural language ut- terance to structural and executable SQL query. We introduce a algorithmic framework named Semantic-Enriched SQL generator (SE-SQL) that enables flexibly access database than rigid API in the application while keeping the per- formance quality for the most commonly used cases. The qualitative result shows that the pro- posed model achieves 88.3% execution accu- racy on WikiSQL task, outperforming baseline by 13% error reduction. Moreover, the frame- work considers several service-oriented needs including low-complexity inference, out-of- table rejection, and text normalization. 1 Introduction The relational database stores a vast of information then support applications in various areas. API and query language normally enforce access to this data. To help retrieving information from database based on utterance, one conventional solution is applying Natural Language Understanding (NLU) model firstly to extract entities as attributes to call the downstream APIs. The extracted entities fulfill the required slot in the pre-defined query templates to retrieve information from the database. Such a rule-based mechanism ensures the input value at runtime while limiting the information retrieval in two major aspects. First, the entity extraction might be constrained under closed domains. It is challenging to apply one trained NLU model when the database is modified, e.g., the new schema or new table. Second, creating or modifying the query template requires numerous human labor and limits the requested query by fixed knowledge of API’s designer. For example, as query the weekly average temperature via date range constraint and aggregation operation, users won’t get the result if there is no such pre-defined query template in the scheme. To address such issues, TEXT2SQL models (Liang and Potts, 2015; Zhong et al., 2017; Xu et al., 2017; Dong and Lapata, 2018; Yu et al., 2019; Dong et al., 2019; Bogin et al., 2019; Lee, 2019; Hwang et al., 2019; Guo and Gao, 2019) are aimed to map natural language utterance to executable SQL query with or without the known database. One example of TEXT2SQL task can be found in Table 1. This paper aims to develop a service-oriented Text-to-SQL parser translating natural language utterance to structural and executable SQL query without the limitation above. We introduce an al- gorithmic framework named Semantic-Enriched SQL generator (SE-SQL) with the following key contributions: Enable flexibly to …

Content truncated to 500 words