RECONSIDERATION ON EVALUATING CORRELATION BY MUTUAL INFORMATION

摘要：

Mutual Information (MI) is the most widely used method to evaluate correlation in statistical natural language processing. However, when reconsidering the essential meaning of MI, it is found out that the amount of shared information is not sufficient to measure the correlation between two random variables; moreover, the symmetry of MI ignores the preference of two meta. Therefore, the paper point out that MI is not sufficient to accurately measure the correlation, and it is more suitable to be used to judge the independency. Furthermore, the author develops the correlation statistics analysis method based on Relative Conditional Entropy (RCE), which could be explained as MI revised by direction. At last, taking new words decision task as an example, the paper applies several statistics methods for comparison. The experiment result proves the correctness of the research result of this paper.

关键词： Mutual Information Relative Conditional Entropy Correlation Preference Model

作者: Wang Daliang Jiang Hongchao Zhang Tao

作者单位: School of Information Engineering,University of Science and Technology,Beijing,China

会议类型: 国际会议

会议名称: 2008年拟人系统国际会议(2008 International Conference on Humanized Systems )(ICHS’08)

会议地点: 北京

会议语种:英文

在线出版日期: 2008-10-18（万方平台首次上网日期，不代表论文的发表时间）