Improving code search with co-attentive representation learning

Searching and reusing existing code from a large-scale codebase,e.g, GitHub, can help developers complete a programming task efficiently. Recently, Gu et al. proposed a deep learning-based model(i.e., DeepCS), which significantly outperformed prior models. TheDeepCS embedded codebase and natural language queries intovectors by two LSTM (long and short-term memory) models separately, and returned developers the code with higher similarityto a code search query. However, such embedding method learnedtwo isolated representations for code and query but ignored theirinternal semantic correlations. As a result, the learned isolated representations of code and query may limit the effectiveness of codesearch.To address the aforementioned issue, we propose a co-attentiverepresentation learning model, i.e., Co-Attentive RepresentationLearning Code Search-CNN (CARLCS-CNN). CARLCS-CNN learnsinterdependent representations for the embedded code and querywith a co-attention mechanism. Generally, such mechanism learnsa correlation matrix between embedded code and query, and coattends their semantic relationship via row/column-wise max-pooling.In this way, the semantic correlation between code and query candirectly affect their individual representations. We evaluate the effectiveness of CARLCS-CNN on Gu et al.'s dataset with 10k queries.Experimental results show that the proposed CARLCS-CNN modelsignificantly outperforms DeepCS by 26.72% in terms of MRR (meanreciprocal rank). Additionally, CARLCS-CNN is five times fasterthan DeepCS in model training and four times in testing.