Research On Deep Learning Based Code Quality Evaluation Method For Academic Papers

The search for the open source code of academic papers is crucial to verify the merits of research results,which directly affects the efficiency and quality of academic research.Therefore,the research on code quality evaluation method for academic papers is of great significance.Deep learning-based code quality evaluation method,which uses feature information of query demand and code as input and relies on deep learning models to evaluate the code quality,has received widespread attention in recent years.The current study has achieved satisfactory results on the quality evaluation of code snippets,but there is still more room for optimization on that of repositories,especially in the relevance,availability and authority.To this end,this dissertation has carried out research on these issues,and the main tasks of the dissertation are as follows: Firstly,three evaluation methods of code quality are constructed.An attention-mechanism based evaluation method is used to measure the relevance between the academic paper and source code according to the paper and readme information.To address the problem that the searched code lacks the description of project implementation,this dissertation proposes an attention-mechanism based method to measure the availability of source code according to the source code and its description information.This dissertation also proposes an influence evaluation method based on the stargazers of code to measure its influence.Secondly,based on the quality evaluation methods of code relevance,code availability and code influence,this dissertation establishes a code search model for academic papers.The model is evaluated on a large-scale dataset containing 10K+ GitHub repositories.The experimental evaluation demonstrates that the code search model proposed in this dissertation achieves better performance in both SR@k and FR@k indicators,and works better than the current code search algorithms of GitHub.The improvements over GitHub in SR@1 and FR@10 are 0.13 and 0.98 respectively,which guarantees the quality of search results.Finally,for the open source code of academic papers,there is still no publicly available dataset at present.In order to promote the research in this field,this dissertation constructs a dataset of correlations between academic papers and open source code in the computer field,with a total of 9225 correlation pairs.In this dissertation,the data of academic paper and open source code is constructed using data collection and document analysis techniques,and the dataset is open source on GitHub for subsequent research.In conclusion,this dissertation implements the code quality evaluation method from the perspective of code relevance,availability and authority,which can provide an effective search scheme and research idea for the code search of academic papers.