查询结果:   张峰逸,彭鑫,陈驰,赵文耘.基于深度学习的代码分析研究综述[J].计算机应用与软件,2018,35(6):9 - 17,22.
中文标题
基于深度学习的代码分析研究综述
发表栏目
综合评述
摘要点击数
858
英文标题
RESEARCH ON CODE ANALYSIS BASED ON DEEP LEARNING
作 者
张峰逸 彭鑫 陈驰 赵文耘 Zhang Fengyi Peng Xin Chen Chi Zhao Wenyun
作者单位
复旦大学软件学院 上海 201203 上海市数据科学重点实验室(复旦大学) 上海 201203    
英文单位
Software School, Fudan University, Shanghai 201203, China Shanghai Key Laboratory of Data Science, Fudan University, Shanghai 201203, China    
关键词
代码分析 深度学习 表征学习
Keywords
Code analysis Deep learning Representation learning
基金项目
作者资料
张峰逸,硕士,主研领域:智能化软件开发。彭鑫,教授。陈驰,博士。赵文耘,教授。 。
文章摘要
随着现代软件规模的不断增大,程序员面临着与日俱增的开发与维护负担,一种辅助他们完成开发流程的工具成为了迫切需求。基于深度学习的代码分析技术从分析代码性质入手,致力于辅助程序员生成或理解代码,成为了软件工程近期研究的热点。总结近期软件工程和人工智能领域中基于深度学习的代码分析研究。论述其一般的技术流程,并从代码表征方法、模型选择以及应用场景三个方面对现有工作进行分类,体现了不同工作的技术特点与设计思路。通过对现有工作的总结整理,认为深度学习在代码分析中的应用还处于初级阶段,既体现了其性能的优越性,又存在着诸如抽象问题描述、数据标注和领域知识理解等问题。在未来的研究中,可能产生突破性进展的发展方向包括建立可标注数据集、设计合理的评判体系以及与知识图谱等新型人工智能技术相结合等。
Abstract
With the continuous increase in the size of modern software, programmers are faced with an ever-increasing burden of development and maintenance. A tool that assists them in completing the development process has become an urgent need. Deep learning based code analysis technology starts with analyzing the nature of code. It is committed to helping programmers generate or understand code, which has become a hot topic in recent software engineering research. This article summarized the recent code analysis research based on deep learning in the field of software engineering and artificial intelligence. It discussed its general technical process, and classified the existing work from three aspects: code representation method, model selection and application scenario, which reflected the technical characteristics and designed ideas of different jobs. By summarizing the existing work, this paper believed that the application of deep learning in the code analysis is still in its infancy, which not only embodies the superiority of its performance, but also has issues such as abstract problem description, data annotation and domain knowledge understanding. In the future research, the development direction that may produce breakthrough progress includes the establishment of a set of data that can be labeled, a rationally designed evaluation system, and a combination of new artificial intelligence technologies such as knowledge maps.
下载PDF全文