Natural Language Interface for Databases Using a Dual-Encoder Model

Proceedings of the 27th International Conference on Computational Linguistics, pages 514–524 Santa Fe, New Mexico, USA, August 20-26, 2018. 514 Natural Language Interface for Databases Using a Dual-Encoder Model Ionel Hosu1, Radu Iacob1, Florin Brad2, Stefan Ruseti1, Traian Rebedea1 1University Politehnica of Bucharest, Romania 2Bitdefender, Bucharest, Romania {ionel.hosu, radu.iacob,stefan.ruseti,traian.rebedea}@cs.pub.ro fbrad@bitdefender.com Abstract We propose a sketch-based two-step neural model for generating structured queries (SQL) based on a user’s request in natural language. The sketch is obtained by using placeholders for specific entities in the SQL query, such as column names, table names, aliases and variables, in a process similar to semantic parsing. The first step is to apply a sequence-to-sequence (SEQ2SEQ) model to determine the most probable SQL sketch based on the request in natural language. Then, a second network designed as a dual-encoder SEQ2SEQ model using both the text query and the previously obtained sketch is employed to generate the final SQL query. Our approach shows improvements over previous approaches on two recent large datasets (WikiSQL and SENLIDB) suitable for data-driven solutions for natural language interfaces for databases. 1 Introduction The quest for a simpler, more intuitive interface to query databases and other sources of structured in- formation is one of the main practical applications of natural language processing. Developing a natural language interface for databases (NLIDB), able to process text queries directly, would be the optimal solution with regards to ease of use, requiring no additional knowledge and allowing non-technical per- sons to interact with the data directly. Nevertheless, even after half of decade of research, this is merely a desideratum, as no current solution is able to perform efficiently for complex databases. From a computational linguistics perspective, natural language interfaces for databases could be treated as a special case of machine translation. Given a text in natural language to query a data source, a NLIDB system needs to output a statement in an artificial language, designed especially by computer scientists to query the database. Several such query languages exist, the most widely used being SQL (Structured Query Language) for structured databases and SPARQL for knowledge bases and ontologies. Traditional NLIDB solutions have been using mainly a combination of rules and heuristics built upon syntactic dependencies and semantic parsing, with machine learning playing only a marginal role. The lack of large datasets suitable for training data-driven solutions is probably the main reason for the lack of complex machine learning approaches. However, a couple of large datasets (Brad et al., 2017; Zhong et al., 2017) have been recently proposed. These corpora allowed deep learning solutions to be deployed. An additional obstacle that needs to be overcome by natural language interfaces for databases, es- pecially by data-driven approaches, is the strong dependence of the corpora on the schema of a given database. Thus, even if one model is trained on a large dataset for a given database schema, it will not have good performance on different schemas. As it is impossible to develop a new dataset with pairs of text queries …

Content truncated to 500 words