Divide-and-conquer Heterogeneous Structure Learning for Text-to-SQL

Anonymous

Abstract: Existing leading Text-to-SQL approaches utilize a unified learning process for semantic and node-edge structural information. However, the unified learning process leads to two major limitations: (i) The mixing of semantic and structural information may cause incorrect linking in structure learning. (ii) The indiscriminate processing of the node graph and the edge graph will cause the loss of the unique property of each graph. In order to address these limitations, we propose a divide-and-conquer Heterogeneous Structure Learning(DCHL) framework for Text-to-SQL, which abstracts the structural information and divides out the semantic information from the original input. Specifically, our framework is featured with the Abstract Graph Construction and Abstract Graph Encoder for the node and edge respectively. We also devise a Semantic-structural Aggregation Mechanism to fuse the divided semantic information and the structure information of nodes and edges. Extensive experiments on three benchmark datasets show that DCHL clearly outperforms strong competitors and achieves new state-of-the-art results. the proposed DCHL achieves competitive results (62.9\% with GLOVE, 72.1\% with ELECTRA) on the cross-domain text-to-SQL benchmark Spider at the time of writing.

Paper Type: long

Research Area: Generation