基于注意力机制和残差连接的BiLSTM-CNN文本分类

BiLSTM-CNN Text Classification Based on Attention Mechanism and Residual Connection

针对文本分类中卷积神经网络(CNN)无法获取文本全局特征、双向循环神经网络(BiLSTM)无法聚焦文本局部特征的问题,提出一种将CNN网络和BiLSTM网络进行融合的模型。该模型引入注意力机制,解决文本分类中无法关注重点单词特征的问题;同时通过引入残差连接,解决网络模型在堆叠多层时出现的神经网络退化问题。实验结果表明,所提出的模型能够更准确地获取文本特征信息,提高文本分类的准确率。

In the text classification, the convolutional neural network(CNN)can not obtain the global features of the text, and the bidirectional cyclic neural network(BiLSTM)can not obtain the local features of the text. A model for fusing a CNN network and a BiLSTM network is proposed. The model introduces the attention mechanism, which solves the problem that the key word features cannot be paid attention to in the text classification. At the same time, by introducing the residual connection, the gradient disappearance of the network model when stacking multiple layers and the parameter update stagnation in the high-level network are solved. The experimental results show that the proposed model can obtain text feature information more accurately and improve the accuracy of text classification.