Advancing Source Code Summaries via Abstract Syntax Trees and Neural Machine Translation

Abstract

Source code summarization generates brief natural language descriptions of code, aiding developer documentation. Traditional neural models often overlook the hierarchical structure of code, leading to inaccuracies. This study introduces the ast-attendgru model, utilizing Abstract Syntax Trees (AST) to capture structural features. Evaluated on Java methods using the BLEU metric, our model outperforms traditional methods and an NLP baseline, especially in cases with minimal documentation. Incorporating AST significantly improves the coherence and accuracy of code summaries.