基于主题模型的文本情感分析研究

发布时间：2018-07-07 23:24

本文选题：文本情感分析 + 主题情感混合模型　；参考：《太原理工大学》2017年硕士论文

【摘要】：互联网中各类评论信息数量巨大、变化迅速,海量的评论表达着信息发布者的情感倾向,这些信息对于政府部门的舆情监控、企业的经营决策和个人的购买决定,都起着至关重要的作用。但通过人工收集分析这些数据不仅成本较高,也无法满足时效性要求,因此文本情感分析受到了学术界与工业界的广泛关注。主题模型是能够从大规模离散数据集中自动提取隐含语义信息的生成概率模型。近年来,主题模型在文本分类、图像分类、热点事件监测、多文档自动文摘、推荐系统等领域得到了广泛应用。主题情感混合模型以主题模型为基础,可以无监督地提取语料的主题信息和对应的情感倾向,在文本情感分析领域得到了越来越多的关注。本文针对长文本和短文本的不同特性,分别提出了词加权LDA算法(Weighted Latent Dirichlet Allocation Algorithm,WLDA)和BJSTM模型(Biterm Joint Sentiment Topic Model,BJSTM)。针对长文本,本文针对现有主题/情感联合分析方法中主题间区分度较低的问题提出了WLDA算法。通过在吉布斯采样中对不同词汇赋予不同权重,增强了具有情感倾向的词汇在采样过程中的影响,从而改善了主题间的区分性。实验表明,与JST模型(Joint Sentiment/Topic model,JST)相比,WLDA不仅在采样中迭代速度快,也能够更好地实现主题提取和情感分类。由于短文本具有稀疏性,面向短文本的主题/情感联合分析方法较少,本文针对这一问题提出适用于短文本的BJSTM模型。该方法为BTM模型(Biterm Topic Model,BTM)增加情感层,对每个双词的情感和主题进行采样。BJSTM模型充分利用了语料级别的丰富词汇共现和词频信息,从某种程度上降低了短文本特征稀疏对主题/情感联合分析的影响。实验表明,与JST和SSTM(Short-text sentiment-topic model,SSTM)相比,BJSTM模型在情感分类和主题关键词提取方面都取得了良好效果。
[Abstract]:There are a lot of comments on the Internet, which are changing rapidly and expressing the emotional tendency of the information publishers. These comments monitor the public opinion of the government, the business decisions of the enterprises and the purchase decisions of the individuals. All play a vital role. However, it is not only expensive to collect and analyze these data manually, but also can not meet the requirements of timeliness. Therefore, text emotional analysis has been widely concerned by academia and industry. Topic model is a generation probability model which can automatically extract implicit semantic information from large scale discrete data sets. In recent years, topic models have been widely used in the fields of text classification, image classification, hot spot event monitoring, multi-document automatic summarization, recommendation system and so on. Subject emotion mixed model is based on topic model, which can extract subject information and corresponding emotional tendency of corpus without supervision, and has been paid more and more attention in the field of text emotion analysis. According to the different characteristics of long text and short text, this paper presents weighted Latent Dirichlet Allocation algorithm WLDA and BJSTM (BJSTM). For the long text, the WLDA algorithm is proposed to solve the problem of low discrimination among the existing methods of topic / emotion joint analysis. By assigning different weights to different words in Gibbs sampling, the influence of affective predisposition words on the sampling process is enhanced, and the distinction between subjects is improved. Experimental results show that WLDA not only has a faster iterative speed in sampling than JST (Joint topic Model JST), but also can achieve better topic extraction and emotional classification. Due to the sparsity of short text and the lack of subject / emotion joint analysis methods for short text, a BJSTM model for short text is proposed in this paper. This method adds affective layer to BTM (Biterm topic Model BTM), and takes full advantage of the abundant lexical co-occurrence and word frequency information at corpus level. To some extent, it reduces the influence of the sparse text features on the thematic / emotional joint analysis. The experimental results show that the BJSTM model has a good effect on emotion classification and topic keyword extraction compared with JST and SSTM (Short-text sentiment-topic Model SSTM).
【学位授予单位】：太原理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1

【参考文献】