改进的卷积神经网络源代码相似性度量方法

A Source Code Similarity Approach Based on Improved Convolutional Neural Networks

源代码相似性是指不同代码段功能上的相似程度,是软件工程领域一项重要的研究问题.现有的方法主要从文本、结构两方面,利用代码的统计学特征计算相似性,其最大缺点就是无法表达代码的语义特征.为解决此类问题,提出了一种融合统计信息的卷积神经网络(statistics information for code embedding-convolutional neural networks,SICE-CNN)源代码相似性检测方法.该方法首先通过词嵌入对源代码进行信息表示,获取代码的词嵌入向量信息;其次,构建CNN训练模型学习源代码文档的嵌入表示;最后,计算源代码对的余弦相似值.实验表明,该方法和一般词嵌入方法相比提高了一定的性能,能较好地检测源代码的语义相似性.

The source code similarity refers to the functional similarity of different code segments,which touches off important research in the field of software engineering.The existing methods mainly extracted texts and structure features manually from source codes to calculate the similarity based on the statistical information in disregard of the semantic characteristics of codes.To solve this problem,a source code similarity detection method based on the CNN was proposed.First,the source code was represented through word embedding to obtain the vector information of word embedding.Second,the CNN training model was constructed to learn the embedded representation of source code documents.Finally,the cosine similarity value of source code pairs was calculated.Experiments show that,the proposed method can certainly improve the performance with respect to the semantic similarity of source codes.