基于半监督卷积神经网络的文本情感分类

Text Sentiment Classification Based on Semi-Supervised Convolutional Neural Network

针对情感分析需要标注大量高质量的训练集来提高模型的准确率,以及卷积神经网络无法充分利用句子中特有的情感信息,提出了一种基于多通道半监督卷积神经网络的情感分析模型。该模型通过重构卷积神经网络的损失函数以及不改变网络的结构来实现模型的半监督学习。同时,模型针对情感分析任务中特有的情感信息来构建文本输入矩阵以及将句子不同特征信息结合形成不同的网络输入通道,使模型在训练过程中从多方面的特征表示来学习输入句子的情感信息。并在COAE 2014和IMDB数据集上进行试验,验证了所提出方法的有效性。

For the sentiment analysis,a large number of high-quality training sets need to be marked to improve the accuracy of the model,and the convolutional neural network can not make full use of the special emotional information in the sentences.This paper proposes an emotion analysis model based on multi-channel semi-supervised convolutional neural network,which reconstructs the loss function of the convolutional neural network,and realizes semi-supervised learning of the model without changing the network structure.At the same time,the model constructs a text input matrix for specific sentiment information in sentiment analysis task,and combines the different feature information of the sentence into different network input channels,so that the model learns the sentiment information of the input sentence from various aspects of the feature representation during the training process.Finally,experiments are performed on COAE 2014 and IMDB datasets to verify the effectiveness of the proposed method.