Fusion of Contrastive Learning and Heterogeneous Graph for Code Summarization

Aiming at the shortcomings of existing code summary generation methods in dealing with the semantic complexity, structural diversity of code and the fusion of search summary and generation summary, a code summary generation method combining contrast learning and heterogeneous graph is proposed. Firstly, a dense vector retrieval framework based on double encoders is adopted to enhance the retrieval accuracy and semantic diversity of large-scale code data. Secondly, the abstract syntax tree is reconstructed into a heterogeneous graph. By introducing various node and edge types, the model's ability to capture complex semantic relationships between different node and edge types is deepened. Finally, the copy gate mechanism is introduced to realize the dynamic fusion of retrieval abstracts and generated model output in semantic level, which guarantees the semantic accuracy and relevance of generated abstracts. Experimental results confirm the effectiveness of this approach in improving the precision and relevance of generated code summaries.