Dependency-based Multi-granularity Code Retrieval

In contemporary times, the technique of code retrieval, aimed at extracting code fragments or files pertinent to a specific query from a code repository, is widely employed in software engineering. This technique significantly assists developers in locating particular code or resolving distinct code issues. However, a major challenge in code retrieval is the disparity between the source code and query statements, which often results in a substantial semantic gap, proving difficult to reconcile. With the advent of the deep learning era, numerous deep techniques are being explored to manage the swelling libraries of code files. However, conventional deep models typically utilize semantic representation grounded on word sequences, neglecting to consider the unique structural data inherent to the code, leading to less-than-optimal performance outcomes. A profound expression based on code abstract syntax trees parsing is code dependency information, which forms the foundation of code execution and aids in capturing the semantics of code efficiently. Consequently, a dependency-based multi-granularity code retrieval model is proposed in this study. This model deftly integrates dependency data into semantics and parses code into statement-level and token-level granularities, thus exploring the semantic relationship between different code granularities. Experimental outcomes demonstrate that the innovative method of integrating dependencies and extracting multi-granularity semantic data noticeably outperforms the results achieved by the presently prevalent baselines.