基于HAN的双通道复合模型的文本情感分类

Text sentiment classification based on HAN andtwo-channel composite model

针对单一的神经网络模型结构简单、传统的注意力模型无法有效提取层次化的文本特征的问题,提出了一种基于HAN的双通道复合模型的文本情感分类。首先,在一个通道上使用双向门限循环(GRU)神经网络提取序列化信息,并引入层次化注意力网络(HAN)学习序列层次化文本信息。其次,在另一通道中借助分解卷积神经网络(CNN)获取局部文本特征,结合HAN学习深层次特征信息。最后,将两个通道进行融合,丰富特征向量,优化文本情感分类效果,提高模型的准确率。在3组中文数据集上进行多组对比实验,本文模型准确率分别达到92.06%,91.08%,92.71%,证明提出模型比单一通道模型表现更出色,使用层次化注意力网络比传统的注意力网络效果更好。

Aiming at the problem that a single neural network model has a simple structure and traditional attention models cannot effectively extract hierarchical text features,propose a text sentiment classification based on HAN and two-channel composite model.Firstly,a bidirectional gated recurrent unit neural network is used to extract serialized information on a channel,and a hierarchical attention network(HAN)is introduced to learn serial hierarchical text information.Secondly,in the other channel,use the convolutional neural network(CNN)to obtain local text features,and use HAN to learn deep-level feature information.Finally,the two channels are fused to enrich the feature vectors,optimize effect of text sentiment classification and improve the accuracy of the model.Comparative experiments on three Chinese data sets show that the accuracy of the models in this paper is 92.06%,91.08%,and 92.71%,respectively.It proves that the proposed model perform better than single channel model,the effect of the HAN is better.