A Software Bug Fixing Approach Based on Knowledge-Enhanced Large Language Models

Software Bug Fixing is a time-consuming task in software development and maintenance. Despite the success of Large Language Models (LLMs) using in Automatic Program Repair (APR), they still have the limitations of generating patches with low accuracy and explainability. In this paper, we propose a software bug-fixing approach based on knowledge-enhanced large language models. First, we collect bugs as well as their fix information from bug tracking systems, such as Github and Stack Overflow. Then, we extract bug entities and inter-entity relationships using Named Entity Recognition (NER) to construct a Bug Knowledge Graph (BKG). Finally, we utilize LLMs (e.g., GPT-4) which is enhanced by the knowledge of the similar historical bugs as well as fix information from BKG to generate patches for new bugs. The experimental results show that the our approach can fix 28.52% (85/298) bugs correctly, which is significantly better than the state-of-the-art approaches. Furthermore, the generated patches are explainable and more credible.