融合知识的主题模型研究及应用

发布时间：2018-03-24 15:57

本文选题：主题模型　切入点：先验知识　出处：《南京理工大学》2017年硕士论文

【摘要】：随着互联网信息的爆炸式发展,网络上的资源和数据越来越丰富,使得人们理解这些动态的海量文本变得非常困难。为了应对这种挑战,我们需要从海量文本中抽取出关键概念,以便人们直观、快速地理解和处理,因此主题模型就应运而生。主题模型算法通过对原始文本中词汇的分析,挖掘隐含其中的主题,这些主题之间的关系,以及随时间演变的情况。但是,近年来研究人员发现,这些没有融合任何人类知识的无监督模型往往会导致生成的主题解释性不强,也就是说,无法生成语义连贯的主题。并且,这些传统的主题模型通常需要大量训练数据。为了解决这些问题,本文进行了融合知识的主题模型研究,并对其在微博话题发现上的应用进行了探索:(1)本文设计了一个融合先验知识的主题模型PLTM,通过对主题模型中两个重要概率分布之一——主题-词分布的改进,从人工提供先验知识和自动挖掘两方面扩展了传统主题模型。并且,针对实际应用中的文本往往会以数据流形式出现的特点,对PLTM模型进行了在线扩展,并设计了两种在线方法。(2)在微博热点话题发现的任务中,设计了增量PLTM模型与两层k-均值聚类和层次聚类的混合聚类结合的方法。本文针对微博语料的特点,采取了较为细致的文本预处理方法,将模型关注的数据对象规模大大减小,并且减少了数据噪音的干扰。与此同时,利用融合知识的主题模型方法,有效地解决了微博短文本数据稀疏性的问题;使用两层k-均值和层次聚类的混合聚类算法,可以快速地将微博聚集到相应的话题下。(3)针对上述两部分的研究工作,本文分别在亚马逊评论数据集和微博数据集中进行了实验分析,验证了模型的实用性和有效性,并且设计了一个用户交互的系统,可以直观地体现模型在实际应用中的效果。
[Abstract]:With the explosive development of Internet information, more and more resources and data are available on the Internet, which makes it very difficult for people to understand these dynamic massive texts. We need to extract the key concepts from the massive text so that people can understand and process them intuitively and quickly, so the topic model comes into being. But in recent years, researchers have found that these unsupervised models that do not integrate any human knowledge tend to lead to less explanatory topics generated, that is, These traditional topic models usually require a lot of training data. In order to solve these problems, the topic model of integrating knowledge is studied in this paper. In this paper, we design a topic model PLTMM, which combines prior knowledge, and improves the topic-word distribution, one of the two important probability distributions in the topic model. In this paper, the traditional topic model is extended from the aspects of manual prior knowledge and automatic mining, and the PLTM model is extended online in view of the fact that the text often appears in the form of data stream in practical applications. In the task of hot topic discovery of Weibo, we design an incremental PLTM model combined with mixed clustering of two-layer K-means clustering and hierarchical clustering. In this paper, we aim at the characteristics of Weibo corpus. A detailed text preprocessing method is adopted to reduce the size of the data object concerned by the model and reduce the interference of the data noise. At the same time, the topic model method of integrating knowledge is used. It effectively solves the problem of data sparsity in Weibo's short text. By using the mixed clustering algorithm based on two-layer k-means and hierarchical clustering, we can quickly gather Weibo under the corresponding topic. In this paper, an experimental analysis on Amazon comment dataset and Weibo dataset is carried out to verify the practicability and validity of the model, and a user interaction system is designed, which can directly reflect the effect of the model in practical application.
【学位授予单位】：南京理工大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.1

【相似文献】