A Method of Constructing and Representing Heterogeneous Graph based on Spider

With the widespread application of database systems, the demand for text-to-SQL technology is continuously growing. In this paper, we propose a novel heterogeneous graph construction and representation method based on the dataset Spider. It can be adapted to subsequent encoding by graph neural network models. Unlike existing graph construction methods, our approach employs constituency parsing to define question nodes, ensuring that each question node possesses complete semantic information. Additionally, schema linking is optimized through prompt engineering by large language models to uncover potential associations. Furthermore, we introduce a part-of-speech(POS) attention mechanism to further enhance the feature representation of question nodes. Experimental results demonstrate a 2.9% improvement in exact match accuracy when integrated with the RAT-SQL model, indicating that the proposed graph construction and representation method effectively enhances the performance of text-to-SQL tasks.