注意力机制在评论文本情感分析中的应用研究  

Research on Application of Attention Mechanism in Comment Text Emotional Analysis

作　　者：申静波[1] 李井辉[1] 孙丽娜[1] SHEN Jing-bo;LI Jing-hui;SUN Li-na(School of Computer and Information Technology,North East Petroleum University,Daqing 163318,China)

机构地区：[1]东北石油大学计算机与信息技术学院,黑龙江大庆163318

出　　处：《计算机技术与发展》2020年第7期169-173,共5页Computer Technology and Development

基　　金：国家青年基金(61702093);黑龙江省高等教育教学改革项目(SJGY20170044)。

摘　　要：长短期记忆神经网络(long short term memory,LSTM)在文本情感分类的准确率方面拥有优秀的表现,能够解决基于长文本序列的模型训练过程中的梯度消失和梯度爆炸等问题。针对传统的LSTM分类模型不能突出体现输出的某个词语对分类类别的贡献程度和重要性的现象,在循环神经网络(RNN)变体长短期记忆人工神经网络(LSTM)隐藏层和输出层之间引入注意力机制,其主要目的是在模型做最后的极性分类预测时,在重要的句子成分加上权重,加大了对最后分类的预测概率的影响因素。实验证明LSTM与注意力机制的融合可突出模型产生动态变化的背景向量以表现出不同输入词语对于输出词语分类的重要性,从而有效提高了分类速率和准确率。Long short term memory(LSTM) has an excellent performance in the accuracy of text emotional classification,and can solve the problems of gradient disappearance and gradient explosion in the process of model training based on long text sequences. In view of the phenomenon that the traditional LSTM classification model cannot incarnate the output word’s contribution to the classification category degree and the importance,we introduce the attention mechanism between hidden layer and output layer in long short term memory(LSTM) variated from circulatory neural network(RNN),aiming to add weights to the important sentence components when the model makes the final prediction of polarity classification and increasing the influence factors on the prediction probability of the final classification. Experiment shows that the integration of LSTM and attention mechanism can highlight the background vector of dynamic changes in the model to show the importance of different input words to the output words classification,thus effectively improving the classification rate and accuracy.

关 键 词：评论文本 情感分析 长短期记忆网络 Seq2Seq模型 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]