基于BiLSTM+Self-Attention的多性格微博情感分类

Research on Multi-personality Microblog Sentiment Classification Based on BiLSTM+Self-Attention

微博作为最受欢迎的社交网络平台之一,是人们表达观点和情感的重要途径。性格影响人的情感表达方式。针对现有微博情感分析很少考虑性格因素这一问题,文章提出一种基于BiLSTM+Self-Attention并结合性格因素的微博情感分类模型(P-BiLSTM-SA)。该模型首先根据“大五”人格理论,基于用户性格将微博文本进行性格分组,然后结合BiLSTM模型和自注意力机制训练出各性格分组的基本分类器,最后采用集成学习方法融合基本分类器预测结果,输出最终的情感标签。为了验证自注意力和性格对情感分类的有效性,文章进行了2组对比实验。第1组实验结果表明,在准确率、精确率、召回率和F1这4个评价指标的综合平均表现上,P-BiLSTM-SA与P-LSTM、P-BiLSTM以及BiLSTM-SA相比,分别提高了0.036、0.017、0.025,说明自注意力机制能有效学习到文本关键信息;第2组实验结果表明,在准确率、精确率、召回率和F1这4个评价指标的综合平均表现上,P-BiLSTM-SA与未结合性格因素的BiLSTM-SA相比,提高了0.012,说明性格因素对情感分类具有一定的帮助。

As one of the most popular social network platforms,microblog is an important way for people to express their views and feelings.Psychological research shows that personality influences the way people express their feelings.In view of the problem that personality is rarely considered in sentiment classification of microblogs,this paper proposes a microblog sentiment classification model,P-BiLSTM-SA,based on BiLSTM+self-attention and combining personality factors.According to"Big Five"theory,the model will first group the microblog texts into different personality groups based on users’personality.Then,the BiLSTM model and the self-attention mechanism are combined to train the basic classifiers of each group.Finally,the ensemble learning method is used to fuse the basic classifiers and output the final affective labels.In order to verify the effectiveness of self-attention and personality in sentiment classifica-tion,two groups of comparative experiments were conducted.The results of the first group of experiments show that,based on the comprehensive average performance of the four evaluation indicators of accuracy,precision,recall rate and F1,the P-BiLSTM-SA proposed in this paper improved 0.036,0.017 and 0.025,compared with the model P-LSTM,P-BiLSTM and BiLSTM-SA.It shows that the self-attention mechan-ism can effectively learn the key information of the text.The results of the second group of experiments show that compared with the BiLSTM-SA without personality factors,the accuracy,precision,recall and F1 of the proposed model P-BiLSTM-SA is improved by 0.012 on average,indicating that the combination of personality factors is useful for sentiment classification.