STNS-CSG: Syntax Tree Networks with Self-Attention for Complex SQL Generation

Text-to-SQL tasks that translate natural questions into corresponding SQL queries aim to help users easily query vast amounts of data stored in relational databases. Most existing studies do not require generating complex SQL queries with subqueries or multiple clauses. However, these complex queries exist widely in real life. Syntax tree network model is proposed to generate complex SQL queries. It suffers from the low accuracy of column prediction. In this paper, we propose a novel model, Syntax Tree Networks with Self-attention for Complex SQL Generation, which develop the syntax tree network model in twofold. The one is applying self-attention mechanism to better capture information in questions, the other is a table predictor that can get table information from questions directly. Experimental results show that our method outperforms the previous state-of-the-art model by 1.9% in exact matching accuracy.