Learning Comment Topics from Code

In this paper, we describe the method of learning comment topics against corresponding code fragments in order to generate topics for the source code that does not contain any documentation or comment. Topics are particular distributions over vocabularies that describe or reveal the meaning or intention of the code. The built system might be helpful for searching code techniques as well as for code classification. The major components of our system are: 1. topic models learned from comments that convert comments into topics; 2. code analyzer and encoder that extract information from codes; 3. recurrent neural networks that learn to predict topics from low-level instruction sequences. The results we obtained are somewhat promising, but cannot be stated as very successful. We spare some of our most interesting unexplored ideas in the future work part, which we believe could work well for the aimed task.