Graph-Fusion RESDSQL: Enhancing NL2SQL Robustness with Dynamic Schema Linking and Contextual Integration

In this paper, we tackle the Natural Language to SQL (NL2SQL) task, which aims to convert user queries in natural language into executable SQL statements, enabling non-professionals to easily access database information without needing programming skills. We propose a new model that builds upon the RESDSQL framework by incorporating graph encoding strategies and a fusion layer. The graph encoding dynamically models tables, columns, and questions using multi-layer graph attention, enhancing the schema linking component of RESDSQL. Additionally, the fusion layer interconnects contextual information between skeleton parsing and schema linking, facilitating information sharing and mutual influence. This approach improves robustness when generating complex SQL statements. Experimental results on the Spider dataset demonstrate that our model outperforms RESDSQL, achieving significant improvements in both Execution Accuracy (EX) and Exact-Match Accuracy (EM), particularly when dealing with complex multi-table joins and nested SQL statements. Our contributions lie in integrating graph structure encoding and self-attention mechanisms into the NL2SQL process, thereby enhancing the model's ability to capture relationships between database entities and natural language queries.