SPG-SQL: Semantic Path Guided SQL Generation Model with Gumbel-Enhanced Graph Pruning

Aiming to reduce the learning cost of database operations, Text-to-SQL methods provide a strategy of automatically generating structured query language (SQL), which is an active research direction by integrating natural language processing and database techniques. However, the existing approaches still have disadvantages in complex structure modeling, semantic path selection, and graph structure pruning, which affect the accuracy and robustness of query generation. To address these challenges, we propose a Semantic Path Guided SQL generation model (called SPG-SQL) that integrates heterogeneous graph modeling, path attention mechanisms, and soft pruning techniques to model semantic relationships between natural languages and database schemas in a hierarchical manner. Extensive experiments on the Spider benchmark demonstrate that SPG-SQL outperforms the state-of-the-art models in exact match accuracy, which validates its effectiveness and practicability.