一种面向隐含主题的上下文树核AContextTree-电子与信息学报.PDFVIP

第 32 卷第 11 期 电 子 与 信 息 学 报 Vol.32No.11 2010 年 11 月 Journal of Electronics Information Technology Nov. 2010 一种面向隐含主题的上下文树核 徐 超 周一民 沈 磊 (北京航空航天大学计算机学院 北京 100191) 摘 要：该文针对上下文树核用于文本表示时缺乏语义信息的问题，提出了一种面向隐含主题的上下文树核构造方 法。首先采用隐含狄利克雷分配将文本中的词语映射到隐含主题空间，然后以隐含主题为单位建立上下文树模型， 最后利用模型间的互信息构造上下文树核。该方法以词的语义类别来定义文本的生成模型，解决了基于词的文本建 模时所遇到的统计数据的稀疏性问题。在文本数据集上的聚类实验结果表明，文中提出的上下文树核能够更好地度 量文本间主题的相似性，提高了文本聚类的性能。 关键词：文本聚类；上下文树核；统计语言模型；隐含狄利克雷分配(LDA) 中图分类号：TP391 文献标识码： A 文章编号：1009-5896(2010)11-2695-06 DOI: 10.3724/SP.J.1146.2009.01493 A Context Tree Kernel Based on Latent Semantic Topic Xu Chao Zhou Yi-min Shen Lei (School of Computer, Beihang University, Beijing 100191, China) Abstract: The lack of semantic information is a critical problem of context tree kernel in text representation. A context tree kernel method based on latent topics is proposed. First, words are mapped to latent topic space through Latent Dirichlet Allocation(LDA). Then, context tree models are built using latent topics. Finally, context tree kernel for text is defined through mutual information between the models. In this approach, document generative models are defined using semantic class instead of words, and the issue of statistic data sparse is solved. The clustering experiment results on text data set show, the proposed context tree kernel is a better measure of topic similarity between documents, and the performance of text clustering is greatly improved. Key words: Text clustering; Context tree kernel; Statistical language models; Latent Dirichlet Allocation (LDA) 1 引言