基于子话题分治匹配的新事件检测

新事件检测是话题检测与跟踪领域的一项重要研究,其任务是实时监控新闻报道流并从中识别新话题.现有方法将话题和报道描述为单一结构的特征向量进行匹配,造成子话题间互为噪声并形成错误语义,从而误导新话题的识别.针对这一缺陷,文中提出基于子话题分治匹配的新事件检测方法,将话题和报道划分为不同子话题,根据相关子话题的比例关系和分布关系建立新话题识别模型.实验在TDT4和TDT5中获得显著改进,最小检测错误代价为0.4061,相应漏检率为0.1859.

New Event Detection Based on Division Comparison of Subtopic

New event detection is an important research in the field of topic detection and tracking, and its task is real-time monitoring the stream of news stories and identifying the new topics in it. Current methods match the topics and stories as they are single-structured vectors of terms, which make the subtopics become noises of each other, and these noises often describe wrong semantics, by which the identification of new topics would be misled. In response to this defect, this paper proposes a new event detection method based on division comparison of subtopics, which divided each topic and story into different subtopics and identified new topic basing on the proportion and distribution relations of the relevant subtopics. This method achieves substantial improvement on TDT4 and TDT5, whose minimum cost of detection error is 0.4061 and missing probability is 0.1859.