基于深度学习的中文文本情感分类及其在舆情分析中的应用研究

发布时间：2018-05-26 21:20

本文选题：深度学习 + 网络舆情　；参考：《湘潭大学》2017年硕士论文

【摘要】：随着中国互联网的爆发式发展,庞大的网络社交群体产生了海量的网络舆情,其情感极其容易被传播和感染。而如何通过庞杂的舆情信息来捕捉分析民众的情感趋势,引导正向的舆情传播,保障社会的安定和谐,是一项极为重要的研究课题。中文文本情感分类作为舆情分析的核心环节之一,备受学者关注和研究。由于中文文本数据具有语义多元、语法特殊性、隐寓性表达等诸多特点,加之当前中文文本情感分类方法大多属于浅层学习方法,存在文本表征能力有限、依赖人工抽取样本特征等缺陷,难以获得较高的中文文本情感分类准确率。为此,如何切合中文文本特点,进一步提升中文文本情感分类的性能是当下网络舆情分析领域迫切需要研究的内容。本文基于深度学习方法开展中文文本情感分类研究,并应用于网络舆情的分析,相关研究内容如下:1.基于主题融合的深度学习情感分类。针对传统深度学习情感分类中只采用词特征的局限性,本文将主题特征与深度学习模型相结合,构建两种主题融合的深度学习情感分类模型:TB_LSTM、TCNN。该两种模型能融合主题特征来获得优质的高层文本特征。实验表明,在二元情感分类中,两种模型最高分类准确率分别达到91.1%和91.9%,比LSTM、CNN、RAE等常用深度学习模型的情感分类准确率平均高出2%左右。2.面向增强特征提取的深度学习情感分类。为提升模型的特征提取能力,本文借鉴特征融合思想,将两种深度学习情感分类模型(TB_LSTM和TCNN)提取的高层文本特征进行融合,以此增强模型的文本特征提取能力,并构建了TB_LSTM+TCNN情感分类模型。实验表明,在相同实验数据条件下,TB_LSTM+TCNN情感分类模型的二元情感分类准确率比TB_LSTM和TCNN的分类准确率高出0.8%-1.6%。3.基于深度学习的多维度舆情分析智能化建模。针对中文网络舆情分析的全面性和准确性需求,本文构建了基于深度学习的多维度舆情智能分析模型,该模型能利用基于增强特征提取深度学习情感分类模型实现精准的中文文本情感分类,并结合主题模型和时间序列模型实现多维度的情感分类、多维度的情感走势分析与预测。4.基于深度学习的多维度舆情分析实证研究。为验证文章提出的深度学习多维度舆情分析模型的有效性,论文以“魏则西事件”为实证案例,基于文章提出的多维度舆情智能分析模型实现对“魏则西事件”的深度解析,实现了多维度的舆情情感走势分析、舆情热点追踪、情感演化刻画,并通过与专家结论对比证明了该模型的有效性和实用性。
[Abstract]:With the explosive development of the Internet in China, a large number of online social groups have generated a huge amount of online public opinion, and their emotions are extremely easy to spread and infect. However, how to capture and analyze the emotional trend of the people, guide the spread of positive public opinion, and ensure the social stability and harmony is an extremely important research topic through the mass public opinion information. As one of the core links in the analysis of public opinion, Chinese text emotion classification has attracted the attention and research of scholars. Because Chinese text data has many characteristics, such as semantic diversity, grammatical particularity, implicit expression and so on, in addition, the current Chinese text emotion classification method mostly belongs to the shallow learning method, and the text representation ability is limited. It is difficult to obtain high accuracy of emotion classification of Chinese text due to the defects of artificial sample extraction. Therefore, how to meet the characteristics of Chinese text and further improve the performance of emotional classification of Chinese text is an urgent need to be studied in the field of network public opinion analysis. Based on the deep learning method, this paper carries out the research of Chinese text emotion classification, and applies it to the analysis of network public opinion. The relevant research contents are as follows: 1. Deep learning emotion classification based on topic fusion. In view of the limitation that only words are used in the traditional deep learning affective classification, this paper combines the topic feature with the depth learning model, and constructs two kinds of deep learning emotion classification model: TBLSTM / TCNN. The two models can combine theme features to obtain high-level text features. The experiments show that the highest classification accuracy of the two models is 91.1% and 91.9% respectively, which is about 2% higher than that of LSTMN CNNRAE and other commonly used deep learning models. Deep learning emotion classification for enhanced feature extraction. In order to improve the feature extraction ability of the model, this paper uses the idea of feature fusion for reference, combines the high-level text features extracted by two kinds of deep learning emotion classification models: TBSP LSTM and TCNN, so as to enhance the text feature extraction ability of the model. And constructed TB_LSTM TCNN emotion classification model. The experimental results show that under the same experimental data, the binary emotion classification accuracy of the model is 0.8-1.6.3. higher than that of TB_LSTM and TCNN. Intelligent modeling of multi-dimensional public opinion analysis based on deep learning. Aiming at the demand of comprehensiveness and accuracy of Chinese network public opinion analysis, this paper constructs a multi-dimensional intelligent analysis model of public opinion based on deep learning. This model can use the enhanced feature extraction depth learning emotion classification model to realize the accurate Chinese text emotion classification, and combines the topic model and the time series model to realize the multi-dimension emotion classification, the multi-dimensional emotion trend analysis and the forecast. 4. An empirical study of multi-dimensional public opinion analysis based on deep learning. In order to verify the validity of the multi-dimensional public opinion analysis model proposed in this paper, this paper takes "Wei Zexi incident" as an empirical case, and realizes the depth analysis of "Wei Zexi event" based on the multi-dimensional intelligent analysis model of public opinion proposed in the paper. The multi-dimensional analysis of the trend of public opinion, the hot spot tracking of public opinion and the depiction of emotional evolution are realized. The validity and practicability of the model are proved by comparison with the expert conclusions.
【学位授予单位】：湘潭大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：G206;G254

【参考文献】