摘要
针对动态话题追踪模型高误报率的现象,提出了动态追踪中的误报检测来判断追踪到的相关报道是否误报,进而降低动态模型的误报率。考虑到新报道是否和话题相关,除了依据两者的相似度外,还涉及时间距离、差值关系、分布关系、追踪到的报道和话题核心报道的相似度四方面内容,给出了误报检测因子计算式。实验采用TDT4测试集合和DET曲线进行评测,通过反复实验获得了误报检测因子δ的阈值,与基于信念网络的动态话题追踪模型相比,使用误报检测后模型的最优(Cdet)norm降低了5.032%。
For the reason of high false alarm probability in dynamic topic tracking, this paper proposed error detection to judge whether the tracked related story was belong to false alarm and then decreased false alarm probability in dynamic topic tracking. Considering a new story was whether related to a topic, not only depending on the similarity between the two, but also depending on difference relationship, distribution relationship, and the similarity between new story and seminal story of existed topic, gave the computation formula of error detection. TDT4 corpora and DET curves were used to run experiments. This paper firstly obtained the threshold of error detection factor δ, the tracking performance of dynamic topic model based on belief network decreases （Cdet）norm by 5. 032% when uses error detection.
出处
《计算机应用研究》
CSCD
北大核心
2015年第2期547-551,共5页
Application Research of Computers
基金
中国博士后科学基金资助项目(20070420700)
河北省自然科学基金资助项目(F2011201146)
关键词
动态话题模型
话题追踪
误报检测
信念网络
dynamic topic model
topic tracking
error detection
belief network
