切片循环神经网络和胶囊网络的性别欺凌识别

Gender Bullying Identification of Sliced Recurrent Neural Networks and Capsule Network

构建基于语境的中文性别欺凌语料库,提出了一种基于语境、结合注意力机制的切片循环神经网络和胶囊网络并联联合算法模型(CASC)。相比传统基于词法规则、句法分析等深度学习神经网络,上述方法可以获取时序词级、句子级、段落级等多个层级高级信息和全局语义信息。同时通过分析上下文语境,挖掘欺凌词之间的依赖关系和深层语义特征,来提高特征表征能力。实验结果表明,上述方法用于网络性别欺凌文本识别精确率为95.33%,召回率为95.83%,衡量模型整体性能的F值为95.58%,准确率为98.78%。从而证明上述方法用于识别性别欺凌文本的有效性。

This paper constructs a context-based Chinese gender bullying corpus.Based on context and attention mechanism,a sliced recurrent neural network and a capsule network parallel joint algorithm model(CASC)is proposed.Compared with the traditional deep learning neural network which is based on lexical rules and syntactic analysis,CASC can obtain multiple levels of advanced information and global semantic information such as sequential word level,sentence level and paragraph level.By analyzing the context and mining the dependence and deep semantic features among bullying words,the ability of feature representation is improved.The experimental results show that the precision rate of CASC for gender bullying text recognition is 95.33%,the recall rate 95.83%,the F value of the overall performance of the model 95.58%,and the accuracy rate 98.78%.These figures prove that CASC is effective to identify the cyber gender bullying text.