Research on natural language query to SQL method with fused table structure

Methods for converting natural language into structured query statements (Natural Language to SQL, NL2SQL) often focus solely on linguistic aspects and table content, neglecting crucial data table structure information. This paper presents an NL2SQL approach that incorporates table structure information. Initially, the method involves associating the question with the corresponding data table column name to acquire input data for the pre-training model BERT (Bidirectional Encoder Representation from Transformers) and obtain feature representations of the text. Subsequently, external vectors representing the structural information of the data table are constructed based on the question's corresponding data table structure information. Finally, the external vectors are fused with the feature vectors produced by the pre-trained model to enhance feature extraction for different subtasks, ultimately improving the accuracy of SQL statement generation. The final results demonstrate that this method significantly enhances the accuracy of the generated SQL statements. Specifically, on the Table QA test set, 90.48% execution accuracy was achieved.