Abstract
Semantic code retrieval is the task of retrieving relevant codes based on natural language queries. Although it is related to other information retrieval tasks, it needs to bridge the gaps between the language used in the code (which is usually syntax-specific and logic-specific) and the natural language which is more suitable for describing ambiguous concepts and ideas. Existing approaches study code retrieval in a natural language for a specific programming language, however it is unwieldy and often requires a large amount of corpus for each language when dealing with multilingual scenarios.Using knowledge distillation of six existing monolingual Teacher Models to train one Student Model—MPLCS (Multi-Programming Language Code Search), this paper proposed a method to support multi-programing language code search tasks. MPLCS has the ability to incorporate multiple languages into one model with low corpus requirements. MPLCS can study the commonality between different programming languages and improve the recall accuracy for small dataset code languages. As for Ruby used in this paper, MPLCS improved its MRR score by 20 to 25%. In addition, MPLCS can compensate the low recall accuracy of monolingual models when perform language retrieval work on other programming languages. And in some cases, MPLCS’ recall accuracy can even outperform the recall accuracy of monolingual models when they perform language retrieval work on themselves.
Highlights
The research on code retrieval can be divided into two broad categories according to the methods used: Information Retrieval-Based Methods and Deep Learning Model-Based
Information Retrieval-Based Methods are more based on traditional search methods, the main idea is to improve work based on text similarity, and to perform code retrieval through search techniques combined with code features
We have prepared a test set for each language and tested it on each monolingual model and Multi-Programming Language Code Search (MPLCS) respectively, the Mean Reciprocal Rank (MRR) results and SuccessRate@k results are shown in Tables 2 and 3 respectively
Summary
The research on code retrieval can be divided into two broad categories according to the methods used: Information Retrieval-Based Methods and Deep Learning Model-Based. Vector Representations is a method to learn the connection between two heterogeneous structures, which maps data of two different structures into the same high-dimensional space [20], so that the corresponding data fall as close as possible to each other in the high-dimensional space, while making the non-corresponding data as far away from each other as possible. Such an approach makes the query process more intuitive, and when performing a search, it is only necessary to find some points in the high-dimensional space that are closest to the target point, i.e., the nearest neighbor problem in the highdimensional space. The code for bubble sort and “bubble sort” is mapped to relatively close locations, as is the case for “quick sort”
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
