Abstract: The quality of service representation vector is the key factor affecting the accuracy of Web service category label recommendation.Most of the existing methods face the problems of incomplete semantic expression and low accuracy when generating service representation vector,which reduces the recommendation accuracy of service category label.Therefore,a Web service category label recommendation method that integrated multi-channel semantic information and attention mechanism was proposed.The embedding of feature words in the service description text was generated using RoBERTa model,and the semantic information extraction channel for feature words of different granularity was established.Then,a global semantic extraction model named FRASRU with fast regular approximate attention mechanism was constructed to realize the fast fusion of feature words' semantic features and global semantic features.Finally,the service representation vector with multi-channel feature fusion was input into the pre-trained sigmiod classifier to realize category label recommendation.Experiments showed that the proposed method was superior to the comparison models and classification methods in the metrics of F1,Recall and Precision,which had good classification performance.

Key words: label recommendation, multi-channel, attention mechanism, Web services

摘要： 服务表征向量的质量是影响Web服务类别标签推荐准确率的关键因素,针对现有方法在生成服务表征向量时普遍存在语义表达不完备和精确度不高,从而影响服务类别标签的推荐准确性的问题,提出一种融合多通道语义信息与注意力机制的Web服务类别标签推荐方法。利用RoBERTa模型生成服务描述文本中特征词的嵌入表示,建立面向不同粒度特征词的语义信息提取通道;构建一种带有快速规则近似注意力机制的全局语义提取模型FRASRU,实现特征词自身语义特征与全局语义特征的快速融合;将多通道特征融合的服务表征向量输入预训练好的sigmiod分类器,实现类别标签推荐。实验表明所提方法优于同类对比模型与分类方法,具有良好的分类效果。

关键词: 标签推荐, 多通道, 注意力机制, Web服务