一种基于改进得分分布的查询项特定阈值方法  

A Term Specific Thresholding Method Based on Improved Score Distribution

作　　者：陆梨花[1] 张连海[1] 

机构地区：[1]中国人民解放军信息工程大学信息系统工程学院,郑州450001

出　　处：《模式识别与人工智能》2015年第5期437-442,共6页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.61175017)资助

摘　　要：为提高语音查询项检索系统的准确率,提出一种基于改进得分分布的查询项特定阈值方法.在系统判决阶段,根据每个查询项的后验得分分布设定不同阈值.后验得分分布用指数混合模型描述,通过无监督的最大期望(EM)算法估计模型参数,最后根据贝叶斯最小风险准则计算阈值.针对EM算法对初始值较为敏感的问题,初始化时采用K-means聚类算法代替随机初始化方法,首先将候选结果得分分为两类,然后计算每类的先验分布并用最大似然法估计模型参数的初始值.实验结果表明该阈值方法有更好的检索性能.To improve the precision of the spoken term detection system, a term specific thresholding method based on improved score distribution is presented. At the decision stage of the system, different thresholds are set for every query according to the posterior scores. The distribution of all posterior scores retrieved for a query term is modeled by exponential mixture model. The parameters are estimated by the expectation maximization （EM） algorithm in an unsupervised manner. The threshold value is calculated by Bayes minimum risk rule. Since EM algorithm is sensitive to initial values, K-means clustering is used in the initialization instead of randomization. Posterior scores are firstly divided into two classes, the prior distributions are calculated and the intial values of the model parameters are estimated by maximum likelihood method. The experimental results show method is better than that of others. that the performance of the proposed thresholding

关 键 词：得分分布 查询项特定阈值 K-MEANS聚类 语音查询项检索 

分 类 号：TN912.3[电子电信—通信与信息系统]