话题跟踪中静态和动态话题模型的核捕捉衰减  

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

作　　者：洪宇[1] 仓玉[1] 姚建民[1] 周国栋[1] 朱巧明[1] 

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006

出　　处：《软件学报》2012年第5期1100-1119,共20页Journal of Software

基　　金：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)(2012AA011102);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)

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)提出的最小检测错误权衡系数评测法,并结合所提出的核捕捉衰减评价方法,对各类主要话题模型进行测试.实验结果显示,结构化的动态话题模型具有最佳的跟踪性能,且突发式增量式学习和时序事件链的更新算法分别给予动态话题模型0.4%和3.3%的性能改进.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.

关 键 词：话题跟踪 静态话题模型 动态话题模型 核捕捉衰减 突发式增量式学习 时序事件链 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]