基于主题的文本情感分类模型研究

发布时间：2018-11-17 13:46

【摘要】：随着WEB2.0技术的普及和电子商务应用的增长,人们更加易于在网站发表自己对于商品的看法和建议。抽取和分析这些情感信息能够利于企业对于商品的改进,同时能够指导用户作出更好的选择。所以,情感分类已经成为了一个研究热点。首先,本文对处理主观性文本信息时所涉及到的一些理论和工具进行研究与探讨,然后基于原有的潜在狄利克雷分布(LDA)模型创建了SO-LDA模型,借助情感语料和分词工具识别出评论文本中的情感词和非情感词,并利用SO-LDA模型进行文本表示,最后利用SVM分类器进行情感倾向性分类。本文所做的工作主要包含以下两个方面：(1)研究情感文本表示模型以及相关技术,提出一种基于LDA的情感主题和其它主题模型。在文本情感倾向性分类之前,首先要做的是针对主观性文本进行建立文档表示模型。因为传统的VSM向量空间模型局限于高维性和稀疏性,所以本文应用了LDA主题模型。论文对LDA进行改进,得到了新的文本表示模型：SO-LDA主题模型。并将其应用到了文本情感倾向性分类领域。(2)分别用LDA和SO-LDA模型解决文本情感分类问题,利用相关的情感语料进行测试,在不同的主题数目对酒店和电脑两个主题进行实验。经实验测试表明,和以往的LDA模型相比,实验设计的SO-LDA模型分类精确度更高。实验中,应用SO-LDA模型对已得到的文本进行建模,将文本中的词语分成两类,情感词和非情感词。根据文本中潜在的情感主题和其它主题对词语进行抽取,然后使用Gibbs抽样算法估计SO-LDA模型的参数,最后进行分类。实验表明,在情感分类问题上,SO-LDA比LDA的情感分类更有效。
[Abstract]:With the popularity of WEB2.0 technology and the growth of e-commerce applications, it is easier for people to publish their opinions and suggestions on goods on the website. Extracting and analyzing these emotional information can help enterprises to improve their products and guide users to make better choices. Therefore, emotional classification has become a research hotspot. First of all, this paper studies and discusses some theories and tools involved in dealing with subjective text information, and then builds a SO-LDA model based on the original potential Delikley distribution (LDA) model. Emotional and non-emotional words are identified by means of affective corpus and word segmentation tools, and SO-LDA model is used for text representation. Finally, SVM classifier is used to classify affective tendency. The work of this paper mainly includes the following two aspects: (1) the emotional text representation model and related techniques are studied, and an emotional theme model and other thematic models based on LDA are proposed. Before classification of text emotion orientation, the first thing to do is to build a document representation model for subjective text. Because the traditional VSM vector space model is limited to high dimension and sparsity, the LDA topic model is applied in this paper. In this paper, we improve LDA and get a new text representation model: SO-LDA topic model. It is applied to the field of text affective preference classification. (2) LDA and SO-LDA models are used to solve the problem of text affective classification, and the related affective corpus is used to test. Experiment with both hotel and computer themes in different number of topics. The experimental results show that the experimental SO-LDA model is more accurate than the previous LDA model. In the experiment, the SO-LDA model is used to model the obtained text, and the words in the text are divided into two categories: affective word and non-emotional word. Words are extracted according to the underlying emotional topics and other topics in the text, then the parameters of the SO-LDA model are estimated by Gibbs sampling algorithm, and then classified. Experiments show that SO-LDA is more effective than LDA in emotional classification.
【学位授予单位】：沈阳工业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.1

【参考文献】