改进Tree⁃LSTM网络的情感分析方法

Sentiment analysis method base on Tree⁃LSTM network

在句子级情感分类任务中,针对深度学习模型易受噪声干扰而导致的分类效果差问题,提出一种结合对抗训练和注意力机制(Attetion)改进树形长短时记忆网络(Tree⁃LSTM)的模型。该模型为多层级结构,包括对抗样本的词嵌入层、Tree⁃LSTM层、注意力机制层。模型中词嵌入层即在词向量添加扰动形成对抗样本,将原始样本与对抗样本一起训练模型,对模型进行正则化处理,增强模型的泛化能力,Tree⁃LSTM层可以提取句子结构特征,注意力机制层对Tree⁃LSTM树的节点赋予不同的权重值以区分不同程度的情感词汇,从而改善模型的分类性能。实验结果显示在数据集SST、MR、COAE2014上,提出模型相比传统模型NBSVM、MNB、LSTM准确率明显提高,比未引入对抗样本的Att⁃TLSTM模型准确率提高,有更快的收敛速度和稳定性,证明该方法能有效提高情感分类任务中的分类性能。

In the sentence⁃level emotion classification tasks,the classification effect of the deep learning model is poor because the model is easily disturbed by noise.Therefore,an improved Tree⁃LSTM(long short⁃term memory)model that combines adversarial training and attention mechanism is proposed.The model is composed of a multi⁃level structure,including the word embedding layer for adversarial sample,the Tree⁃LSTM layer and the attention mechanism layer.In the word embedding layer,perturbation is added to the word vector to form the adversarial sample,and then both the original sample and the adversarial sample are used together to train the model.The model is regularized and its generalization ability is enhanced.In the Tree⁃LSTM layer,sentence structure features can be extracted.Attention mechanism layer gives different weights to the nodes of Tree⁃LSTM tree to distinguish emotional vocabularies at different levels,so as to improve the classification performance of the model.The experimental results are shown on the data sets SST,MR and COAE2014.In comparison with the traditional models of NBSVM,MNB and LSTM,the accuracy of the proposed model is obviously improved.In comparison with the Att⁃TLSTM model without introducing the adversarial sample,the accuracy of the proposed model is improved.In addition,its convergence speed is faster and more stable,which proves that this method can effectively improve the classification performance in emotion classification tasks.