Codex comment: lexical and syntactic enhanced code comment generation

Abstract

Insufficient or poor-quality code comments are the result of inexperience or intentional omission, which can ‎hinder software comprehension and pose challenges, particularly in the context of maintenance within a ‎team-programming environment. Here, the smart tools for a more detailed understanding, such as ‎automatic generation of comprehensible code comments, can give insights about functionality and purpose of ‎every element of the code. During the last few years, studies were conducted about this problem since code ‎comments are the only way for human constructively to code is to be conveyed and are therefore highly ‎important to developers. The most popular solutions are inspired by recent breakthroughs in NMT in casting ‎the problem of generating code comments as an NMT problem consisting of translating code into sounding-‎the-best-of-human comments. With such a level of advancement, we will be introducing Codex Comment, the ‎latest and more advanced deep learning-based approach, to automatically generate high-quality comments ‎for any code. The following is based on the Transformer model first proposed by Google researchers, which ‎achieved impressive results in solving diverse sequence-to-sequence tasks to achieve faithful code comment ‎generation. Codex Comment generates hybrid codes, which carry both lexical as well as syntactic information, ‎hence a more structure-sensitive representation of source code compared to others in sub-word-based ‎approaches. As in the codex model, it generates the harmonic information from its tokens of operations‎.