话题跟踪中静态和动态话题模型的核捕捉衰减软件学报ISSN1000-9825,CODENRU_UEWE-mail:jos@JournalofSoftware,____,23(5):11001119[doi:10.3724/SP.J.1001.____.04045]__169;中国科学院软件研究所版权所有.Tel/Fa_:+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)DescendingKernelTrackofStaticandDynamicTopicModelsinTopicTrackingHONGYu+,CANGYu,YAOJian-Min,ZHOUGuo-Dong,ZHUQiao-Ming(SchoolofComputerScienceandTechnology,SoochowUniversity,Suzhou215006,China)+Correspondingauthor:E-mail:hongy@HongY,CangY,YaoJM,ZhouGD,ZhuQM.Descendingkerneltrackofstaticanddynamictopicmodelsintopictracking.JournalofSoftware,____,23(5):11001119./1000-9825/4045.htmAbstract:Topictrackingisataskinresearchonidentifying,miningandself-organizingrelevantinformationtonewstopics.Itskeyissueistoestablishstatisticalmodelsthatadaptthekindofnewstopic.Thisincludestwoaspects:oneistopicalstructure;theotheristopicevolution.Thispaperfocusesoncomparingandanalyzingthefeaturesofthreemainkindsoftopicmodelsincludingwordsbag,hierarchicaltreeandchain.Differentperformancesofstaticanddynamictopicmodelsaredeeplydiscussed,andatermoverlappingratebasedevaluationmethod,namelydescendingkerneltrack,isproposedtoevaluatetheabilitiesofstaticanddynamictopicmodelsontrackingthetrendoftopicdevelopment.Onthisbasis,thispaperrespectivelyproposestwomethodsofburstbasedincrementallearningandtemporaleventchaintoimprovetheperformanceofcapturingtopickernelsofdynamictopicmodels.E_perimentsadopttheinternational-standardcorpusTDT4andminimumdetectionerrortradeoffevaluationmethodproposedbyNIST(NationalInstituteofStandardsandTechnology),alongwithdescendingkerneltrackmethodtoevaluatethemaintopicmodels.Theresultsshowthatstructuraldynamicmodelshavethebesttrackingperformance,andtheburstbasedincrementallearningalgorithmandtemporaleventchainachieve0.4%and3.3%improvementrespectively.Keywords:topictracking;statictopicmodel;dynamictopicmodel;descendingkerneltrack;bustyfeaturebasedincrementallearning;temporaleventchain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(NationalInstituteofStandardsandTechnology)提出的最小基金项目:国家自然科学基金(61003152,60970057,60873105,909____4,60970056);国家高技术研究发展计划(863)(____AA收稿时间:____-04-26;修改时间:____-12-15;定稿时间:____-04-28011102);国家教育部博士点基金(____3____10006);苏州市应用基础研究计划基金(SYG____30)
