【作者】 赵斌；

【导师】 丁宇新；

【作者基本信息】 哈尔滨工业大学 ， 计算机科学与技术， 2013， 硕士

【摘要】 当今世界，随着信息技术的不断进步，各类信息资源都以较快的速度增长并呈现出海量的特征，其中数据主要还是以文本的形式出现。信息量已经满足人们的需求，但是如何高效地管理并且使用海量数据成为目前迫切需要解决的问题，这促进了对文本分类相关领域的研究。文本分类技术研究的主要核心内容包括两个部分：分类模型和文本表示。目前文本表示方法可以分为两种类型，一种类型是引入语言学特征；另一种类型利用统计学方法挖掘出文本的主题信息。前者由于需要较为复杂的语言学特征处理从而降低了整个系统的效率，其实用性受到影响；后者的典型代表就是PLSA语义模型以及LDA语义模型。语义模型是一种基于统计学理论的概率模型，模型构建出“文档-主题-词汇”三层结构来得到数据集中的潜在语义（主题）。本文给出了基于点间互信息的LDA（Point-wise Mutual Information LatentDirichlet Allocation，PMI-LDA）主题模型和基于最近距离的LS（Laplace Score）主题选择算法。LDA模型不是判别模型，而是一个生成模型，在生成文本的过程中通过EM算法得到潜在的主题层。但是，在利用LDA主题模型生成文本的过程中，模型会同等对待文本中的每一个单词，这样会造成主题向高频词倾斜，同时还会造成主题重叠现象。本文的主要贡献有三点：首先，提出PMI-LDA主题模型，模型能够克服主题向高频词倾斜及主题重叠的问题，使得我们提取出的文本主题更能表征一篇文本。实验证明，本课题提出的算法是可行的。然后，本文还从两个角度来评价提取出主题的优劣，一是从主题一致性、可读性上来评价，从主题本身出发，根据主题所包含的单词的可读性和一致性，对主题的优劣进行评价；二是从模型中主题的区别能力和相似性上来评价，从整个模型出发，根据主题的相似性和区别能力来评价提取出的主题的优劣。从实验中可以清晰的看到，PMI-LDA主题模型提取出的主题，不管是在可读性、一致性上，还是在区别能力、相似性上都优于LDA主题模型提取的主题。最后，当使用主题作为文本的特征时，根据主题存在优劣的事实，不同的主题不应该同一而视，课题提出了基于最近距离的LS算法来计算主题的权值，并应用到文本分类中。更多还原

【Abstract】 In today’s world, with the continuous advancement of information technology,the internet has become the most widely used in the world, the most informativeinformation library. Meanwhile, the various types of information resources at afaster rate growth and showing the characteristics of the mass, which data is mainlyin the form of text appear. The amount of information has to meet the needs ofpeople, but how to efficiently manage and use the vast amounts of data as there is anurgent need to address the problem, which has contributed to the study of the relatedareas of text classification. Text classification technology research core contentconsists of two parts: classification model and text representation. The method ofthe text representation can be divided into two types; one type is to introducelinguistic features to improve the performance of the text classification. Anothertype is to use of statistical methods to dig out the theme of the text information. Theformer due to the need for more complex linguistic characteristics of treatmentwhich reduces the efficiency of the entire system, so that it’s practical affected. Atypical representative of the latter is the PLSA semantic model as well as the LDAsemantic model. The semantic model is a probabilistic model, based in the statisticaltheory to model the “document-theme-word” latent semantic dataset.In this paper, we propose Latent Dirichlet Allocation based on the Point-wiseMutual Information and the Laplace Score theme selection algorithms based on thenearest distance. LDA model is not a discriminate model, but a generative model.Potential theme layer will be got in the process of generating text via the EMalgorithm. However, in the process of using the original LDA algorithm generatestexts; the algorithm will treat every word in the text as equally important, this willcause the theme to the high-frequency words tilt, while also causing the thematicoverlap. The main contribution of this paper is three things: First, propose a LDAmodel based on point-wise mutual information, the model is able to solve theproblem of topic tilting to the high-frequency words and themes overlap, allows usto extract the text theme which can better characterize a text. The experiments showthan the algorithm of this topic is feasible. Second, the article also extracted fromtwo perspectives to evaluate the theme. One is according to the readability andconsistency of the topics included the words, the other is according to the distinctiveand independence of the topic from the model. Can clearly see from the experiment,the LDA topic model based on point-wise mutual information is better than theoriginal LDA topic model according to the readability, consistency, or distinction, independence of the theme. Third, a theme exist advantages and disadvantages, asthe characteristics of the text when the topic vector, each dimension of the vectorshould be not same, subject propose LS algorithm based on the shortest distance tocalculate the theme of the right value which been applied to text application.更多还原