基于事件-时间关联模型的话题跟踪研究

Research on topic tracking based on event-time relation model

针对话题跟踪的任务是从时序新闻报道流中实时识别和挖掘相关于特定新闻话题的报道,本文提出一种事件-时间关联模型(EventTime Relation Model,ETRM)用来展开话题跟踪研究。ETRM将相关报道的时间属性引入向量空间模型,话题跟踪过程中将话题与相关报道相同特征项的时间相关度应用于相关性判定机制,同时基于时间的分布属性调整特征向量的权重分配,实现话题模型的自适应学习更新。实验采用DET曲线评测系统性能,结果显示相比于传统的话题模型,ETRM能够更加准确地追踪到话题焦点演化趋势,有效提高了话题跟踪系统的性能。

This paper proposes an Event-Time Relation Model（ abbr. ETRM） to study topic tracking for its task that is to identify and mining subsequent on-topic stories in the temporal story stream. The ETRM introduces the time property of the story to the vector space model,apply time correlations of same feature to the correlation decision mechanism in topic tracking process,adjusting feature vector weight allocation based on time property to implement subject model of adaptive learning at the same time. Experiment adopts DET curve performance evaluation system,the results show that ETRM can more accurately track the topic focus of evolution trend compared with the traditional model of subject,effectively improve the performance of topic tracking system.