基于sigmoid函数的评分特征规整在计算机辅助学习中的应用  

Application of Scoring Features Adjustment in CALL Based on Sigmoid Function

作　　者：严可[1] 蒋海曦[2,3] 

机构地区：[1]中国科学技术大学讯飞语音实验室,合肥230027 [2]四川大学经济学院 [3]四川师范大学成都学院外语系,成都610065

出　　处：《成都纺织高等专科学校学报》2011年第3期40-46,共7页Journal of Chengdu Textile College

摘　　要：计算机辅助语言学习是近十年来日益热门的研究课题,语音评分是其重要的组成部分。目前,国内大部分评分系统在运用机器评分特征预测人工分时,都是采用简单的线性回归模型。虽然该模型具有很好的集外推广性,但要求机器的评分特征与人的主观评分呈近似的线性关系,这一点在实际情况中往往难以满足。对此,本文引入sigmoid函数对评分特征进行规整,并采用数据驱动的方式得到其参数,使得提取的评分特征更符合评分员的主观评分准则。同时,在算法优化的问题上,本文将该方法完全地嵌入经典的线性回归模型中求解,大大提高了其收敛速度。实验表明,在普通话水平考试的单字、词语朗读及中学生考试翻译题自动评分方面,该算法使得评分的系统性能得到明显的优化。Computer Assisted Language Learning （CALL） has been a hot issue m last oecade ana pronunciation scoring is one of its important parts. At present, most scoring systems adopt simple linear regression models by using machine scoring features to predict human score. Such model can be generalized, but the linear relationship between machine scoring features and human subjective scoring is difficult to meet. Therefore, sigmoid function was introduced to adjust the scoring features and made them more consistent with human scoring criteria by adopting digital - driven method. Meanwhile, in order to optimize algorithm, such method was completely embedded into classical linear regression model to improve its convergence speed. Experiments showed that the performance of scoring system by this method had been optimized obviously in such aspects as the pronunciation and reading of Chinese Mandarin proficiency test and middle school students＇ translation test.

关 键 词：语音评测 SIGMOID函数 评分特征 计算机辅助语言学习 普通话水平测试 

分 类 号：G642.475[文化科学—高等教育学] O29[文化科学—教育学]