Research on Cross-Domain Text2SQL Enhancing Database Schema Awareness

Text2SQL has become an important research direction of natural language processing because of its abundant practical scenarios and research value. RAT-SQL is a Text2SQL method based on relational self-attention, which solves the domain generalization problem by improving the schema encoding, schema linking and feature representations in the encoder. However, the approach only focuses on adding the known relation representation into self-attention to achieve alignment between entities and schema items, which cannot well fuse external heterogeneous information. The ability of the parser to generalize to invisible target domains remains limited. To address this problem, this paper proposes a new approach for Text2SQL tasks that injects factual relations within domain knowledge and extends some of the capabilities of RAT-SQL by jointly encoding natural language questions, database schemas and exogenous knowledge. Experiments show that the accuracy of EDSA-SQL is significantly improved compared to RAT-SQL on the publicly available dataset Spider.