一种融合外部特征的改进主题模型  

An Improved Topic Model Integrating Extra-Features

作　　者：杨如意[1] 刘东苏[1] 李慧[1] 

机构地区：[1]西安电子科技大学经济与管理学院,西安710126

出　　处：《现代图书情报技术》2016年第1期48-54,共7页New Technology of Library and Information Service

基　　金：国家自然科学基金青年基金项目"基于可信语义wiki的知识库构建方法与应用研究"(项目编号:71203173)的研究成果之一

摘　　要：【目的】在LDA模型基础上融合时间和作者特征,提出动态作者主题(DAT)模型,更好地揭示文本内容、主题和作者之间的关系。【应用背景】从海量文本中实现特征抽取和语义挖掘已经成为情报研究人员的重要工作。【方法】获取NIPS会议论文作为数据集并进行预处理,按发表年份划分到每个时间片形成一阶马尔科夫链,使用困惑度确定最优主题数,并在每个时间片内通过吉布斯采样估算作者主题概率分布和主题词项概率分布。【结果】实验结果表明,该模型将文档表示为作者主题概率分布和主题词项概率分布,时间维度上可观测主题强度变化和作者兴趣变化。【结论】DAT模型能够有效地融合文档内容与外部特征,实现文本挖掘。[Objective] In order to reveal the relationships between contents, topics and authors of documents, this paper presents the Dynamic Author Topic （DAT） model which extends LDA model. [Context] Extracting features from large-scale texts is an important job for informatics researchers. [Methods] Firstly, collect the NIPS conference papers as data set and make preprocessing with them. Then divide data set into parts by published time, which forms a first-order Markov-chain. Then use perplexity to ensure the number of topics. At last, use Gibbs sampling to estimate the author-topic and topic-words distributions in each time slice. [Results] The results of experiments show that the document is represented as probability distributions of topics-words and authors-topics. On the dimension of time, the revolution of authors and topics can be observed. [Conclusions] DAT model can integrate contents and extra-features efficiently and accomplish text mining.

关 键 词：LDA模型 DAT模型 文本挖掘 吉布斯采样 

分 类 号：G254[文化科学—图书馆学]