两种Zipf定律拟合方法的对比实证研究  

Comparative Empirical Study on Zipf's Law with two Fitting Methods

作　　者：路高飞[1] 韩普[1] 沈思[1] 

机构地区：[1]南京大学信息管理学院,南京210093

出　　处：《图书情报工作》2012年第24期71-76,126,共7页Library and Information Service

基　　金：国家社会科学基金一般项目"以创新和质量为导向的哲学社会科学评价体系实证研究"(项目编号:12BTQ037)研究成果之一

摘　　要：采用普通最小二乘法和最大似然估计两种方法分别对Zipf定律的分布进行拟合比较研究。普通最小二乘法是目前曲线拟合中普遍采用的方法,而最大似然估计是曲线拟合更为合理的方法。分别利用三组中文和三组英文语料,对两种方法进行对比实证研究,结果发现最大似然估计方法的拟合更合理,对Zipf定律的拟合比普通最小二乘法好,同时发现英文词汇比中文更好地符合Zipf定律分布,即中文语料不严格符合经典的Zipf定律分布。本研究可以为幂律拟合的研究提供一些参考。Respectively with the methods of Ordinary Least Square and Maximum Likelihood Estimation, this paper does a fitting comparative study on the distribution of Zipf＇ s law. Ordinary Least Square is a popular method in curvefitting, while the Maximum Likelihood is a more reasonable approach. Based on the three Chinese corpuses and three English eorpuses, this paper empirically compares the two fitting methods. The results show that Maximum Likelihood Estimation is much better than Ordinary Least Square at calculating the slope of Zips＇ s law; English corpus more accorded with the Zipf＇ s law distribution than Chinese ones, which means Chinese corpus doesn＇ t strictly conform to the classic Zipf＇ s law distribution. This paper provides a reference for the power - law fitting research.

关 键 词：Zipf定律 普通最小二乘法 最大似然估计 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]