COMBINE: A Pipeline for SQL Generation from Natural Language

Abstract

Accessing data stored in relational databases requires an understanding of the database schema and mainly a query language such as SQL, which, while powerful, is difficult to master. In this sense, recent researches try to approach systems to facilitate this task, in particular by making Text-to-SQL models that attempt to map a question in Natural Language (NL) to the corresponding SQL query. In this paper, we present COMBINE, a pipeline for SQL generation from NL, in which we combine two existing models, RATSQL (We used the version RATSQL v3+BERT; paper’s url: arxiv.org/abs/1911.04942.) and BRIDGE (We used the version BRIDGE v1+BERT; paper’s url: aclweb.org/anthology/2020.findings-emnlp.438/.), that are based on recent advances in Deep Learning (DL) for Natural Language Processing (NLP). Our model is evaluated on the Spider challenge, using Exact Matching Accuracy (EMA) and Execution Accuracy (EA) metrics. Our experimental evaluation demonstrates that COMBINE outperforms the two used models in the same challenge, and at the time of writing, achieving the state of the art in EA with 70%, and competitive result in EMA with 71.4%, on Spider Dev Set.