基于跨层连接的多通道DBiSAC网络欺凌检测模型

A DBiSAC cyberbullying detection model based on cross-layer connection

目的:网络欺凌文本存在特征稀疏、用词不规范、语义模糊等问题,导致简单的神经网络无法充分提取其语义特征和句法特征。因此,提出了一种基于跨层连接的多通道DBiSAC网络欺凌检测模型。方法:先利用Glove预训练模型实现文本的词向量化表示。其次构建并行的多尺度深度可分离卷积(MSDSC)和采用了跨层连接策略的双向简单循环(CBiSRU)的多通道特征提取网络,分别提取文本的局部特征和提取文本的上下文中语义信息的全局特征。再将局部特征和全局特征拼接之后利用多头注意力机制(MHA)对其重要特征信息进行捕捉。然后利用胶囊网络(Capsnet)对文本序列中空间层次的语义特征进行提取。最后将提取的文本特征信息通过含有Softmax函数的全连接层分类器输出分类结果。结果:实验表明,提出的模型在两个网络欺凌的数据集上均取得了最优的结果。结论:本文模型在能够有效提高模型准确率。。

Aims:The problems such as sparse features,irregular wording,and fuzzy semantics in cyberbullying text led to the fact that a simple neural network could not fully extract its semantic and syntactic features.A multi-channel DBiSAC network bullying detection model based on cross-layer connection was proposed.Methods:The Glove pre-training model was used to realize the word vector representation.A parallel multi-scale deep separable convolution(MSDSC)and a bi-directional simple loop(CBiSRU)multi-channel feature extraction network with cross-layer connection strategy were constructed to extract the local features of the text and the global features of the semantic information in the context of the text respectively.After the local feature and the global feature were spliced,the important feature information was captured by the multi-head attention mechanism(MHA).Then the capsule network(Capsnet)was used to extract the spatial level features in the text sequence.The extracted text feature information was passed through the full connection layer classifier with Softmax function to output the classification results.Results:The proposed model achieved the best results on two cyberbullying data sets.Conclusions:The model can effectively improve the accuracy.