Abstract: This paper analyses the Internet news reports and finds their characteristics such as redundancy, low centralization of the discussions and the topic drift. An Online Topic Detection(ODT) method for Internet is proposed. It defines the sub-topic to ignore the redundancies reports, presents the double-lays configuration for the low centralization of the discussions, and advances a topic tracking algorithm based on the sliding window. A topic detection system is build according to the method. The system is tested by the real data from the Internet. The results present that this method is better than the single-pass method for ODT. The CDet of the method is under 0.14.

Key words: Online Topic Detection(ODT), Topic Detection and Tracking(TDT), text clustering

摘要：

针对互联网新闻报道冗余多、议题发散、易漂移等特点，提出一种面向互联网的在线话题检测算法。该算法针对冗余问题提出子话题概念，针对议题发散问题建立双层检测结构，针对话题漂移问题提出基于滑动窗口的跟踪策略。应用该算法建立网上话题检测系统，通过来源于互联网的真实数据进行测试。结果表明，算法性能优于传统的单路径聚类算法，其最小错误代价率低于0.14。

关键词: 在线话题检测, 话题检测与跟踪, 文本聚类

CLC Number: 

• TP311