Mixed Node and Edge GraphSAGE Based Schema Encoding for Text-to-SQL

In text-to-SQL tasks, it is important to construct a model that is applicable in databases of different domains, and how to improve the generalization and induction capabilities of the model is a pivot challenge to be addressed. Traditional approaches are limited by the following shortcomings: 1) limited to transductive models, which do not perform as well as inductive models when generating embeddings for unseen nodes of graphs in cross-domain scenarios; 2) insufficient attention to the importance of structural information in relational databases and do not distinguish local and non-local relations. Therefore, we propose a Mixed Node and Edge Graph SAmple and aggreGatE text-to-SQL model, MNEGraphSAGE for short, which is based on graph sampling and propagation algorithm for node and edge features and converges faster. We give full attention to graph network edge features by mixing the iterative process of edge features and node features in parallel. We also adopt different aggregation methods for edge features of different distances. In the Spider dataset covering multiple domains, our method achieves 64.02% without the help of pre-trained models, outperforming other models of the same size.