Code Similarity via Natural Language Descriptions

Meital Ben-Sinai, M.Sc. Thesis Seminar
Wednesday, 18.3.2015, 12:30
Taub 601
Prof. E. Yahav

Code similarity is a central challenge in many programming related applications, such as code search, automatic translation, and programming education. We present a novel approach for establishing the similarity of code fragments by computing textual similarity between their corresponding textual descriptions. In order to find textual descriptions for code fragments, we leverage the collective knowledge captured in question answering sites, blog posts and other sources. Because our notion of code similarity is based on similarity of corresponding textual descriptions, our approach can determine semantic relatedness and similarity of code across different libraries and even across different programming languages, a task considered extremely difficult using traditional approaches. We have implemented our approach using data obtained from the popular Q&A site, Stackoverflow, and used it to determine the similarity of 100,000 pairs of code fragments. To evaluate our approach, we implemented a crowdsourcing system that allows users to label the similarity of code fragments, and used it to build a massive corpus of 6500 labeled program pairs. Our results show that our technique is effective in determining similarity, and gains high precision, recall and accuracy.

Back to the index of events