GADESQL: Graph Attention Diffusion Enhanced Text-To-SQL with Single and Multi-hop Relations

GADESQL: Graph Attention Diffusion Enhanced Text-To-SQL with Single and Multi-hop Relations

연구 분야: Databases

학회: International Conference on Web Information Systems Engineering

Text-To-SQL is crucial for enabling users without technical expertise to effectively extract important information from databases. The graph-based encoder has been successfully employed in this field. However, existing methods often adopt a node-centric approach, focusing on single-hop edge relations. This approach gives rise to two main issues: 1) failure to differentiate between single-hop and multi-hop relations among nodes; 2) ignoring the valuable multi-hop reasoning information between nodes. To tackle these challenges, we propose a Graph Attention Diffusion Enhanced Text-To-SQL(GADESQL) model that enables multi-hop reasoning among nodes. With GAD, information can propagate efficiently through multi-hop paths, uniquely integrating single-hop and multi-hop relations during the graph iteration process. Furthermore, we employ Semantic Dependency Parsing for natural language analysis, constructing a semantic analysis tree for questions to enhance the effective connection between question tokens and Schema structures. Experiments on the cross-domain dataset Spider demonstrate that our model possesses strong generalization capabilities, achieving certain performance improvements over existing works.