A Heterogeneous Graph to Abstract Syntax Tree Framework for Text-to-SQL.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Database Management

Background:

• Text-to-SQL systems convert natural language queries into SQL.
• Existing models struggle with complex database schemas and SQL structure.
• Higher-order semantics in graph encoding and permutation biases in AST construction are limitations.

Purpose of the Study:

• To propose a generic heterogeneous graph to abstract syntax tree (HG2AST) framework for improved Text-to-SQL.
• To integrate structural knowledge into statistical models for enhanced performance.
• To address limitations of traditional encoder-decoder models in Text-to-SQL.

Main Methods:

• Developed a line graph enhanced encoder (LGESQL) for iterative node and edge feature updates via dual graph message passing.
• Implemented a grammar-based decoder to construct SQL abstract syntax trees (ASTs) followed by post-processing.
• Introduced a golden tree-oriented learning (GTL) algorithm to mitigate permutation biases during AST node expansion.

Main Results:

• The proposed HG2AST framework demonstrates superior performance across various Text-to-SQL datasets.
• The LGESQL encoder effectively captures higher-order semantics in heterogeneous graph encoding.
• The GTL algorithm successfully controls AST node expansion, reducing permutation biases.

Conclusions:

• The HG2AST framework offers a significant advancement in Text-to-SQL accuracy and applicability.
• Integrating structural knowledge through graph and tree models is crucial for complex NLP tasks.
• The approach shows broad applicability in single/multi-table, single/cross-domain, and multilingual settings.