Accelerating Code Search with Deep Hashing and Code Classification

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2534 - 2544 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Accelerating Code Search with Deep Hashing and Code Classification Wenchao Gu1∗, Yanlin Wang2,†, Lun Du2, Hongyu Zhang3, Shi Han2, Dongmei Zhang2, and Michael R. Lyu1 1 Department of Computer Science and Engineering, The Chinese University of Hong Kong, China. 2 Microsoft Research Asia, Beijing, China 3 The University of Newcastle, Australia Abstract Code search is to search reusable code snippets from source code corpus based on natural lan- guages queries. Deep learning-based methods on code search have shown promising results. However, previous methods focus on retrieval accuracy, but lacked attention to the efficiency of the retrieval process. We propose a novel method CoSHC to accelerate code search with deep hashing and code classification, aiming to perform efficient code search without sacri- ficing too much accuracy. To evaluate the ef- fectiveness of CoSHC, we apply our method on five code search models. Extensive experi- mental results indicate that compared with pre- vious code search baselines, CoSHC can save more than 90% of retrieval time meanwhile preserving at least 99% of retrieval accuracy. 1 Introduction Code reuse is a common practice during software development process. It improves programming productivity as developers’ time and energy can be saved by reusing existing code. According to pre- vious studies (Brandt et al., 2009; Lv et al., 2015), many developers tend to use natural language to describe the functionality of desired code snippets and search the Internet/code corpus for code reuse. Many code search approaches (Brandt et al., 2009; McMillan et al., 2011; Lv et al., 2015; Du et al., 2021) have been proposed over the years. With the rapid growth of open source code bases and the development of deep learning technology, recently deep learning based approaches have be- come popular for tackling the code search prob- lem (Gu et al., 2018; Husain et al., 2019; Gu et al., 2021). Some of these approaches adopt neural network models to encode source code and query descriptions into representation vectors in the same ∗ Work done while this author was an intern at Microsoft Research. Wenchao Gu (wcgu@cse.cuhk.edu.hk). † Yanlin Wang is the corresponding author (yanl- wang@microsoft.com). embedding space. The distance between the repre- sentation vectors whose original code or descrip- tion are semantically similar should be small. Other approaches (Feng et al., 2020; Guo et al., 2021; Du et al., 2021) regard the code search task as a binary classification task, and calculate the probability of code matching the query. In the past, deep learning-based methods focused on retrieval accuracy, but lacked attention to the efficiency of retrieval on large-scale code corpus. However, both types of these deep learning-based approaches directly rank all the source code snip- pets in the corpus during searching, which will incur a large amount of computational cost. For the approaches that separately encode code and de- scription representation vectors, the similarity of the …

Content truncated to 500 words