基于seq2seq模型的标签推荐方法

针对NPM平台上存在大量的软件包没有标签或标记不完善的问题, 提出一种基于seq2seq模型的深度学习方法为软件包推荐标签. 首先, 利用ECMAScript工具分析软件包的源码构建出包的函数调用图, 遍历函数调用图从而将软件包转换成一组具有包语义信息的函数调用序列; 其次, 训练seq2seq模型, 并将训练好的模型用于软件包的标签推荐工作, 该模型能将包的函数调用序列映射到一组预测的标签序列上, 从而完成软件包的标签推荐. 实验结果表明, 该方法能为软件包推荐一组合理的标签, 准确率达82.6%.

Tag Recommendation Mehtod Based on seq2seq Model

Aiming at the problem that a large number of software packages had no tags or imperfect tags on node package manager (NPM) platform, we proposed a deep learning method based on the seq2seq model to recommend tags for software packages. Firstly, we used ECMAScript tools to analyze the source code of software package, constructed function call graphs of the package, and traversed the function call graph, so as to convert the software package into a set of function call sequences with package semantic information. Secondly, we trained the seq2seq model and applied the trained model to tag recommendation of software package. The trained model could map the function call sequence of package to a group of predicted tag sequence, so as to complete the tag recommendation of software packge. The experimental results show that the method can recommend a reasonable set of tags for the software package, and the accuracy is 82.6%.