HeSQLNet: A Heterogeneous graph neural network for SQL-to-Text generation

Context:

Understanding the semantics of SQL queries is crucial for maintaining code and reusing functionalities in database access and management. However, SQL queries often remain challenging to comprehend, even for expert users. In this work, we address this challenge by focusing on SQL-to-Text, a task that translates SQL queries into corresponding natural language questions. Existing approaches predominantly encode SQL queries using their Abstract Syntax Tree (AST) representation and then decode this structure into textual explanations. However, these methods often treat the AST as a homogeneous graph, overlooking the diverse relationships between its nodes, such as parent–child and sibling relationships.

Objective:

To address this issue, this paper introduces HeSQLNet: a Heterogeneous Graph Neural Network for SQL-to-Text Generation.

Methods:

Specifically, we first propose a Heterogeneous Feature Graph (HFG), which augments the AST with six distinct edge types to better capture the heterogeneous relationships inherent in SQL queries. We further develop a heterogeneous graph neural network with attention, leveraging a two-stage aggregation process to effectively extract and encode these heterogeneous features within the HFG. The enriched HFG representation is then incorporated into an encoder–decoder framework, called HeSQLNet, to generate natural language descriptions of SQL queries. To assess the ability of SQL-to-Text models to handle complex queries and demonstrate compositional generalization, we introduce SpiderComGen, a new compositional generalization dataset derived from the Spider dataset.

Results:

We conduct extensive experiments on both the widely-used and our proposed datasets. The experimental results reveal that HeSQLNet significantly outperforms existing state-of-the-art approaches in both effectiveness and generalization capability. Additionally, compared to the recent large language models, human evaluations and case studies show that HeSQLNet delivers not only accurate results but also more concise outputs.

Conclusion:

Our HeSQLNet proves that heterogeneous feature fusion and extraction significantly improve SQL-to-Text generation.