FCSO: Source Code Summarization by Fusing Multiple Code Features and Ensuring Self-consistency Output

Source code summarization is the process of generating a concise and generalized natural language summary from a given source code, which can facilitate software developers to comprehend and use the code better. Currently, most research on source code summarization generation focuses on either converting the source code into abstract syntax tree (AST) sequences or directly converting it into code segments and then feeding these representations into deep learning models. However, these single representation approaches ignore the semantic features of source code and destroy the structure of the abstract syntax tree, which affects the quality of the generated source code summarization. In this paper, we propose a novel source code summarization approach that fuses multiple code features into self-consistency output (FCSO). Our approach is based on a graph neural network encoder and a CodeBERT encoder with a self-attention mechanism. It extracts the sentence feature attention vector and the AST feature attention vector of the source code for feature fusion. Then, it inputs them into the Transformer decoder. Furthermore, to generate more accurate source code summaries, we adopt a new decoding strategy called self-consistency. It samples different inference paths, uses a penalty mechanism to calculate their similarity scores, and ultimately selects the most consistent answer. Our experimental results demonstrate that our proposed approach outperforms standard baseline approaches. On the Python dataset, the BLEU score, METEOR score, and ROUGE_L score increase by 11.13 \(\%\) , 9.12 \(\%\) , and 7.88 \(\%\) , respectively. These results show that our approach provides a promising direction for future research on source code summarization.