一种基于统计流形学习的文本分类算法

发布时间：2018-03-23 17:11

本文选题：文本分类　切入点：流形学习　出处：《中国科学技术大学》2017年硕士论文

【摘要】：文本是一种常见的数据形式,人们每天都会使用文本这一信息载体与他人进行交流,互联网中每时每刻都会产生海量的文本数据。文本分类在信息检索、数据挖掘、情感分析等任务上都发挥着巨大的作用。根据提取特征方式的不同,文本分类算法可以分为以下三大类:基于统计的文本分类算法,基于语义相似度的文本分类算法以及基于深度学习的文本分类算法。常见的基于统计的文本分类算法有词频-逆向文件频率模型(TF-IDF),朴素贝叶斯等等。这些方法把词作为特征项,词出现的次数作为权值,然后把文本表示为一个特征向量,最后使用分类器进行文本分类。这些方法假设相似的文本中有很多相同的词,但是这忽略了不同词之间的语义相似性。基于语义相似度的文本分类方法通常根据文本的主题信息对文本的相似度进行度量,如主题模型等等,但是这些方法不能够清晰地捕捉到词和文本的主题多样性。近年来,深度学习方法吸引了许多研究者的注意,但是这些方法,比如卷积神经网络或者循环神经网络等等,也有着一些不足。比如梯度消失问题以及大规模参数训练带来的时间消耗。本文提出了一种基于统计流形学习的文本分类算法,它提供了一个基于潜在主题分布的文本概率模型表示。该模型假设同一个主题下的词服从高斯分布,然后文本被表示为一个混合高斯模型,最后利用统计流形学习的方法可以对文本之间的距离进行度量。本文的主要工作包括:(1)从文本的生成过程出发,提出了一种文本表示的概率模型。每个主题被表示为一个高斯分布,文本被表示为一个高斯混合模型。这种概率模型可以对文本和词的主题多样性进行较好的描述。(2)通过使用概率模型对文本上的主题分布进行描述,文本建模的计算时间复杂度降低为O(n),n是文本中单词数量。主题模型中训练速度和语料依赖性的问题得到了改良。(3)通过统计流形学习方法,对文本概率模型的距离进行了度量,提供了一种度量概率模型的新思路。(4)在实验部分,通过三组不同任务的实验,验证了本文所提算法的有效性以及高斯混合模型描述混合主题下词向量分布的能力。
[Abstract]:Text is a common data form. People use text to communicate with others every day. Text categorization is used in information retrieval and data mining. According to the different ways of extracting features, text classification algorithms can be divided into the following three categories: statistical based text classification algorithm, Text classification algorithm based on semantic similarity and text classification algorithm based on in-depth learning. Common statistical text classification algorithms include word frequency reverse file frequency model TF-IDFU, naive Bayes and so on. The number of occurrences of a word is used as a weight, then the text is represented as a feature vector, and finally a classifier is used to classify the text. These methods assume that there are many identical words in similar text. But this ignores the semantic similarity between different words. Text classification methods based on semantic similarity usually measure the text similarity according to the subject information of the text, such as topic model, etc. However, these methods can not clearly capture the diversity of words and texts. In recent years, in-depth learning methods have attracted the attention of many researchers, but these methods, such as convolution neural networks or cyclic neural networks, etc. For example, the gradient vanishing problem and the time consumption caused by large-scale parameter training. In this paper, a text classification algorithm based on statistical manifold learning is proposed. It provides a representation of the text probabilistic model based on the distribution of potential topics, which assumes that the words under the same theme are distributed from Gao Si, and then the text is represented as a mixed Gao Si model. Finally, the distance between texts can be measured by using the method of statistical manifold learning. A probabilistic model for text representation is proposed. Each topic is represented as a Gao Si distribution. The text is represented as a Gao Si mixed model. This probability model can describe the topic diversity of the text and word better by using the probabilistic model to describe the topic distribution on the text. The computational time complexity of text modeling is reduced to the number of words in the text. The problem of training speed and corpus dependence in the topic model is improved. The distance of the text probability model is measured by using the statistical manifold learning method. This paper provides a new way to measure the probability model. In the experiment part, the validity of the proposed algorithm and the ability of Gao Si hybrid model to describe the word vector distribution under the mixed theme are verified by three groups of experiments with different tasks.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1

【相似文献】