基于并行双向门控循环单元与自注意力机制的中文文本情感分类

在基于深度学习的文本情感分类研究领域中，目前传统的模型主要是序列结构，即采用单一的预训练词向量来表示文本从而作为神经网络的输入，然而使用某一种预训练的词向量会存在未登录词和词语语义学习不充分的问题。针对此问题，提出基于并行双向门控循环单元（gated recurrent unit，GRU）网络与自注意力机制的文本情感分类模型，利用两种词向量对文本进行表示并作为并行双向GRU网络的输入，通过上下两个通道分别对文本进行上下文信息的捕捉，得到表征向量，再依靠自注意力机制学习词语权重并加权，最后对两个通道的输出向量进行向量融合，作为输入进入全连接层判别情感倾向。将本文模型与多个传统模型在两个公共数据集上进行实验验证，结果表明本文模型在查准率、查全率、F₁值和准确率等性能指标上相比于双向门控循环单元网络模型、双向长短时记忆网络模型和双向门控循环单元网络与自注意力机制的单通道网络模型均有所提升。

Chinese text sentiment classification based on parallel bi-directional gated recurrent unit and self-attention

Current traditional deep learning models mainly employ a sequence structure in the field of text sentiment classification, and a single pre-trained word vector is used to represent the text as an input to the neural network. However, the use of a certain pre-trained word vector gives rise to problems in that the unregistered words and the semantic learning of the words are insufficient. To solve these problems, we propose a text sentiment classification model based on a parallel bi-directional gated recurrent unit (GRU) network and self-attention. Two word vectors are used to represent the text as the input to the bi-directional GRU network, and the context information is captured by the two channels. The attention mechanism is then used to learn the weight of the words from the representation vector. Finally, the two channels' output vectors are merged, and used as the input into the fully connected layer to discriminate the sentiment tendency. Our proposed model was compared with several traditional models using two different datasets. The results show that our model has better performance indicators, including precision, recall, F₁ and accuracy.