基于改进LDA-CNN-BiLSTM模型的社交媒体情感分析研究

Research on Social Media Sentiment Analysis Based on Improved LDA-CNN-BiLSTM Model

针对社交媒体情感分析忽略情感特征的长距离语义关系,无法精确捕获带有情感色彩的特征词,过度依赖人工标注等问题,本文提出一种改进LDA-CNN-BiLSTM模型,旨在实现对微博舆情事件的情感分析研究。首先,通过对微博舆情事件评论文本进行数据采集和数据预处理,获取“喜悦”“愤怒”和“哀伤”三种类别情感文本。其次,构建融合LDA模型、情感词典和人工标注的算法并用于情感特征词提取,使用Word2Vec将经过特征提取后的情感文本转换为词向量。最后,构建CNN-BiLSTM模型,利用卷积神经网络提取文本的关键特征,长短时记忆网络捕获长距离语义特征,从而完成情感分类任务。实验结果表明,本文方法的精确率、召回率、F1值和准确率分别为0.8946、0.8841、0.8893和0.8778,整体实验结果均优于现有的机器学习和深度学习模型,并且融合LDA模型和情感词典的实验结果均有明显提升,其F1值比实验中的六种机器学习模型平均提升3.66%,比七种深度学习模型平均提升1.84%。综上,本文方法能够应用于社交媒体的情感分析任务,并有效感知舆情事件的情感态势,具有一定的研究价值。

To solve the problems of social media sentiment analysis,such as ignoring long-distance semantic relationships of emotional features,failing to capture feature words with emotional color accurately,and over-relying on manual annotation.This paper proposes an improved LDA-CNN-BILSTM model to realize sentiment analysis and research on microblog public opinion events.Firstly,three categories of emotional texts of joy,anger,and sadness were obtained through data collection and data preprocessing of comments on Weibo public opinion events.Secondly,an algorithm combining the LDA model,emotion dictionary,and manual annotation is constructed to extract emotion feature words.Then,word2Vec is used to convert emotion text after feature extraction into a word vector.Finally,the CNN-BilSTM model was constructed,and the convolutional neural network was used to extract the key features of the text.The long and short time memory network was used to capture the long-distance semantic features to complete the emotion classification task.Experimental results show that the proposed method’s precision,recall,F1-score,and accuracy are 0.8946,0.8841,0.8893,and 0.8778,respectively.The overall experimental results are better than existing machine learning and deep learning models,and the experimental results of the LDA model and emotion dictionary are significantly improved.Its F1-score is 3.66%higher than the six machine learning models in the experiment and 1.84%higher than that of the seven deep learning models.To sum up,the method in this paper can be applied to the emotion analysis task of social media and effectively perceive the emotional situation of public opinion events,which has outstanding research value.