GCB-AST: A Software Defect Prediction Method Based on GraphCodeBert Feature Extraction

software quality assurance faces significant challenges due to the escalating complexity and scale of modern software systems. “Traditional software defect prediction methods rely on software metrics but suffer from limited feature extraction capabilities, failing to fully characterize code defect patterns and thus compromising prediction accuracy. This paper aims to propose a new approach to software defect prediction to address the current challenges in the field. This paper proposes a software defect prediction method GMB-AST, which combined the semantic feature extraction method of GraphCodeBert and the Abstract Syntax Tree (AST) of code, and verified the effectiveness of GMB-AST in software defect prediction through experiments. Experimental results on Promise Datasets demonstrate statistically significant improvements over the state-of-the-art DP-CCL model, with absolute gains of 3.9% in recall, 2.5% in precision, and 5.9% in F1-score (p<0.01). By bridging semantic understanding and structural analysis, this work enables more effective automated quality assurance for large-scale software systems.