基于特征融合深度学习网络的情感分析模型

Sentiment Analysis Model Based on Feature Fusion Deep Learning Network

现如今,深度学习技术迅速发展,在情感分析任务中被大量使用。针对传统神经网络模型中,卷积神经网络难以捕捉文本前后文关联语义信息以及长短时记忆网络训练所需时间长、缺乏深层次信息提取能力的问题,提出了一种双通道卷积神经网络和循环网络变体的特征融合情感分析模型(CSRMA),该模型能够获取更全面的情感特征,加快训练速度。该模型使用GloVe进行词向量化,将词向量分别传入卷积神经网络通道,和引入注意力机制与最大池化的BiSRU通道,得到局部深层次情感语义和前后文时序情感语义,最后融合特征进行分类,输出情感极性,完成情感分析任务。论文主要在四个英文短文本情感数据集上,与传统神经网络模型进行对照实验,从实验结果看来,CSRMA模型对于情感语料的分类准确性有进一步提高,训练时耗费时间缩短,具有良好的泛化性。

Nowadays,deep learning technology develops rapidly and is widely used in emotional analysis tasks.In the traditional neural network model,convolutional neural network is difficult to capture the semantic information associated with the text before and after,and long short-term memory network training takes a long time and lacks the deep information extraction ability.In this paper,a two-channel hybrid model based on convolution neural network and recurrent neural network variants is proposed,which can obtain more comprehensive emotional characteristics and speed up training.In this model,GloVe is used to vectoring words,and the word vector is respectively introduced into convolutional neural network channel,and BiSRU channel with maximum pooling and attention mechanism,so as to obtain the local deep emotional semantics and the previous and subsequent sequential emotional semantics.Finally,the fusion features are classified,the emotional polarity is output,and the emotional analysis task is completed.The CSRAM model is compared with the traditional neural network model on four English short text emotional datasets.From the experimental results,the CSRAM model can further improve the classification accuracy of emotional corpus,reduce the training time,and have good generalization.