Code Similarity via Natural Language Descriptions

Code similarity is a central challenge in many programming related applications, such as code search, automatic translation, and plagiarism detection. In this work, we reduce the problem of semantic relatedness between code fragments into a problem of semantic relatedness of textual descriptions. Our main idea is that we can use the relationship between code and its textual descriptions as established in question-answering sites such as STACKOVERFLOW. Consequently, we can determine semantic relatedness and similarity, of code fragments across different programming languages, a task considered extremely difficult using traditional approaches. We have implemented our approach, and used crowed-sourced labeling of similarity to evaluate it over 1500 pairs of code fragments. Results show that we gain around 80% precision and 75% recall, and demonstrate the promise of this approach.