话题跟踪中静态和动态话题模型的核捕捉衰减

软件学报ISSN 1000-9825, CODEN RUXUEW E-mail: jos@

Journal of Software,2012,23(5):1100 1119 [doi: 10.3724/SP.J.1001.2012.04045]

©中国科学院软件研究所版权所有. Tel/Fax: +86-10-62562563

话题跟踪中静态和动态话题模型的核捕捉衰减

洪 宇+, 仓 玉, 姚建民, 周国栋, 朱巧明

(苏州大学 计算机科学与技术学院,江苏 苏州 215006)

Descending Kernel Track of Static and Dynamic Topic Models in Topic Tracking

HONG Yu+, CANG Yu, YAO Jian-Min, ZHOU Guo-Dong, ZHU Qiao-Ming

(School of Computer Science and Technology, Soochow University, Suzhou 215006, China)

+ Corresponding author: E-mail: hongy@

Hong Y, Cang Y, Yao JM, Zhou GD, Zhu QM. Descending kernel track of static and dynamic topic models in

topic tracking. Journal of Software, 2012,23(5):1100 1119. /1000-9825/4045.htm

Abstract: Topic tracking is a task in research on identifying, mining and self-organizing relevant information to

news topics. Its key issue is to establish statistical models that adapt the kind of news topic. This includes two

aspects: one is topical structure; the other is topic evolution. This paper focuses on comparing and analyzing the

features of three main kinds of topic models including words bag, hierarchical tree and chain. Different

performances of static and dynamic topic models are deeply discussed, and a term overlapping rate based evaluation

method, namely descending kernel track, is proposed to evaluate the abilities of static and dynamic topic models on

tracking the trend of topic development. On this basis, this paper respectively proposes two methods of burst based

incremental learning and temporal event chain to improve the performance of capturing topic kernels of dynamic

topic models. Experiments adopt the international-standard corpus TDT4 and minimum detection error tradeoff

evaluation method proposed by NIST (National Institute of Standards and Technology), along with descending

kernel track method to evaluate the main topic models. The results show that structural dynamic models have the

best tracking performance, and the burst based incremental learning algorithm and temporal event chain achieve

0.4% and 3.3% improvement respectively.

Key words: topic tracking; static topic model; dynamic topic model; descending kernel track; busty feature based

incremental learning; temporal event chain

摘 要: 话题跟踪是一项针对新闻话题进行相关信息识别、挖掘和自组织的研究课题,其关键问题之一是如何建

立符合话题形态的统计模型.话题形态的研究涉及两个问题,其一是话题的结构特性,其二是话题变形.对比分析了

现有词包式、层次树式和链式这3类主流话题模型的形态特征,尤其深入探讨了静态和动态话题模型拟合话题脉络

的优势和劣势,并提出一种基于特征重叠比的核捕捉衰减评价策略,专门用于衡量静态和动态话题模型追踪话题发

展趋势的能力.在此基础上,分别给出突发式增量式学习方法和时序事件链的更新算法,借以提高动态话题模型的核

捕捉性能.实验基于国际标准评测语料TDT4,采用NIST(National Institute of Standards and Technology)提出的最小

基金项目: 国家自然科学基金(61003152, 60970057, 60873105, 90920004, 60970056); 国家高技术研究发展计划(863)(2012AA

收稿时间: 2010-04-26; 修改时间: 2010-12-15; 定稿时间: 2011-04-28 011102); 国家教育部博士点基金(200932011 10006); 苏州市应用基础研究计划基金(SYG201030)

搜索“diyifanwen.net”或“第一范文网”即可找到本站免费阅读全部范文。收藏本站方便下次阅读，第一范文网，提供最新人文社科话题跟踪中静态和动态话题模型的核捕捉衰减全文阅读和word下载服务。