基于去相关化的低秩矩阵分解对口语能力的评估方法  

A Low Rank Matrix Decomposition Based on Decorrelationfor Speaking Proficiency Aassessment

作　　者：林思岑 LIN Sicen(Institute of Foreign Languages,Xi’an Medical University,Xi’an 710021,China)

机构地区：[1]西安医学院外国语学院,陕西西安710021

出　　处：《微型电脑应用》2021年第2期33-36,共4页Microcomputer Applications

基　　金：陕西省教育厅专项科研项目(18JK0650);西安医学院配套基金项目(2016PT42)。

摘　　要：口语流利性的自动评分是一个分类问题。训练模型参数,将输入流利性特征映射到对应的真实分数,然后用于预测输入话语的分数。因此,为了估计模型参数以可靠地预测分数,必须提供正确的真实分数作为目标输入。然而,要从人工打分的分数中确定正确的基本事实分数并不简单,因为这些分数包括主观偏见。所以,在此提出了一种基于解相关惩罚低秩矩阵分解的法语口语流利性分数的基本真实值估计方法。该方法将人工评分的得分分解为偏差和潜在得分,最大化皮尔逊相关。最后通过实验证明了该方法的有效性。Automatic scoring of spoken fluency is a classification problem.To solve the problem,we train model parameters,map input fluency features to corresponding true scores,and then use them to predict the score of input utterances.Therefore,in order to estimate the model parameters to reliably predict the score,it is necessary to provide the correct score as the target input.However,it is not easy to determine the basic fact of correct scores from the manually scored scores,because these scores include subjective biases.Therefore,a basic true value estimation method of French oral fluency score based on decorrelation penalty low-rank matrix decomposition is proposed.This method decomposes the scores of manual score into deviations and potential scores to maximize Pearson correlation.Finally,experiments prove the effectiveness of the method.

关 键 词：机器学习 深度神经网络 低秩矩阵分解 口语能力 解相关优化 评估 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]