Generating pseudo-code from source code using deep learning

Authors

Abdulaziz Alhefdhi, University of WollongongFollow
Hoa K. Dam, University of WollongongFollow
Hideaki Hata, Nara Institute of Science & Technology Japan
Aditya K. Ghose, University of WollongongFollow

RIS ID

133234

Publication Details

Alhefdhi, A., Dam, H. Khanh., Hata, H. & Ghose, A. (2018). Generating pseudo-code from source code using deep learning. Proceedings - 25th Australasian Software Engineering Conference, ASWEC 2018 (pp. 21-25). United States: IEEE.

Abstract

Pseudo-code written in natural language and mathematical expressions is a useful description of source code. Pseudocode aids programmers in understanding the code written in a programming language they are not familiar with. However, writing pseudo-code for each code statement is labour intensive. In this paper, we propose a novel approach to automatically generate pseudo-code from source code using Neural Machine Translation. Our model is built upon the deep learning encoderdecoder using the attention-based Long Short-Term Memory architecture to capture the long-term dependencies in both source code and pseudo-code. An empirical evaluation on a real Python dataset demonstrates the applicability of our approach in practice.