Automated source code summarization via transformer

Abstract

Source code summarization is a comprehensible description of a program’s functionality. The code summarization assists developers to understand large portions of source code, thus reducing the time taken to comprehend a program’s capabilities. To automate the code summarization, programs have used RNN-based neural architecture to create neural network models for this natural language translation. However, the RNN-based neural architecture has two particular limitations which are its disability to process the non-sequential structure of the source codes and missing out on the long-term relationships between code tokens. My proposed approach of using Transformer neural architecture is able to overcome these limitations. Compared against the RNN-based neural network models, the Transformer network model has shown significantly better experimental results of BLEU 1, 2, 3 and 4 scores, ranging between three to seven scores higher, METEOR score of three higher and ROUGE-L score of one higher.