CORES: COde REpresentation Summarization for Code Search

Abstract

With the growth of the consumer electronics market, the software development industry is facing new opportunities and an increased focus on code retrieval techniques to improve efficiency and reduce costs. Code search aims to retrieve and reuse code from extensive repositories based on a search query with specific requirements. Recently, pre-trained model-based approaches have become popular because of grasping semantic representations of code snippets and search queries accurately. However, such approaches ignore the inconsistency between code and query statements due to the redundant tokens, such as definitions and punctuation marks in the code snippets, which hinder the matching accuracy. To tackle such disadvantage, in this paper, two strategies are proposed based on explicit or implicit code representation summarization. By summarizing the code representation, the redundancy in the code is removed and the inconsistency between code and query statements is alleviated. For the explicit code representation summarization-based strategy, different views of contextual information are obtained and summarized based on different scales of pyramidal dilated convolution. As to the implicit code representation summarization-based strategy, covariance is directly applied to constrain the code representation to ensure de-redundancy. Experimental results on six benchmark datasets show both strategies outperform the current State-Of-The-Art model CORES by 1.2% on average MRR scores.