Methods of source code summarization: from abstract syntax trees to heterogeneous graph neural networks

Abstract

This article examines methods of automatic source code summarization as a means of generating concise natural language descriptions that reflect program functionality and structure. The main problem addressed is the limited ability of many existing approaches to capture semantic context beyond local syntactic patterns, which reduces the practical usefulness of generated summaries. The study analyzes representative approaches based on abstract syntax trees, graph-based program representations, and heterogeneous graph neural networks, focusing on how different structural models encode program information. The methodology consists of a comparative analysis of code representations and context propagation mechanisms used in modern summarization systems. The results show that syntax-oriented models effectively represent local hierarchical structure but fail to model long-range dependencies, while graph-based approaches provide richer semantic coverage by integrating control-flow and data-flow relations. The analysis further demonstrates that heterogeneous graph models enable multi-level context aggregation, improving summary relevance and contextual completeness. The practical significance of the results lies in identifying representation strategies that are most suitable for real-world applications such as automated documentation, code comprehension tools, and developer assistants. The findings clarify current limitations and outline directions for designing semantically informed and context-aware code summarization systems.