Code Search Method Based on Multimodal Representation

Abstract

Developers tend to search and reuse code snippets from large-scale corpora while implementing some of the features that existed in development. This will improve the efficiency of development. Code search is to search for semantically relevant code snippets based on a given natural language query. In existing methods, the semantic similarity between code and query is quantified as their distance in the shared vector space. To improve the vector space and map the code vector and query vector into a shared vector space so that the semantically similar code-query pairs are close to each other, we propose a code search method with multimodal representations. It can better enhance the semantic relationship between code snippets and queries. Experiments on Java datasets show that the multimodal representation model MulCS improves the quality of code search. MulCS outperforms several existing advanced models in several performance metrics.