基于互信息的话题特征选择方法研究

Research on Topic Feature Selection Based on Mutual Information

特征选择是舆情监测系统构建的关键步骤之一,好的特征选择方法可以在降低系统时间消耗的同时,提高舆情监测的准确性。针对话题特征提取方法 ITF-IDF没有考虑类别信息的缺点,尝试将互信息用于话题特征提取,提出改进的互信息计算方法 CMI和DCMI。CMI方法融入了聚类思想,对新闻报道进行分组验证,DCMI在CM I的基础上,将时间信息量化为时间距离,实现特征互信息的动态更新。实验结果显示,DCM I的性能明显优于基本互信息方法和ITF-IDF方法。

Feature selection is one of the key steps to build an opinion monitoring system,and a good feature selection method should both reduce time consumption and improve the accuracy of opinion monitoring. Since the existing topic feature selection method ITF-IDF doesn＇t consider category information,we attempt to use mutual information for topic feature selection and propose tw o modified mutual information methods CM I and DCM I. CM I merges clustering and testifies new s stories by group; DCM I quantifies time information as time distance to dynamically update feature mutual information. Experiment results indicate that the performance of DCM I is obviously better than the basic mutual information and ITF-IDF methods.