计算机用于英语背诵题的自动评分技术初探  

PRIMARY STUDY ON COMPUTERISED AUTOMATIC MARKING OF ENGLISH RECITATION PROFICIENCY

作　　者：严可[1] 胡国平[2] 魏思[1,2] 李萌涛[3] 杨晓果[3] 冯国栋[3] 

机构地区：[1]中国科学技术大学人机语音通信实验室,安微合肥230027 [2]安徽中科大讯飞信息科技股份有限公司,安微合肥230088 [3]中国科学技术大学英语系,安微合肥230027

出　　处：《计算机应用与软件》2010年第7期164-168,共5页Computer Applications and Software

摘　　要：计算机辅助语言学习日益成为研究重点,但目前主要开展的还是朗读题型的自动评分技术研究。在已有的朗读题自动评分工作基础上,首次提出并尝试背诵题的自动评分研究工作。通过分析背诵题的特点和规律,提出了用"结合静音模型和吸收模型的句子并联网络"进行背诵文本的自动识别,并在识别结果基础上选取更适合背诵题的评分特征,再结合朗读评分中的自适应技术,实现英语背诵题的自动评分。在2006年外研杯新概念英语大赛采集的342份数据上,系统取得了较好的评分性能:人机评分相关度达到0.732,平均分差0.358,达到了专家间相关度的90%,在完整度和流畅度的细节评分上,自动评分水平甚至超过了人工评分水平。Computer-assisted language learning growingly becomes the emphasis in research,but what have been mainly developing now is on automatic marking technique for the exam type of reading proficiency.Based on our previous work in this regard,we extend our work to automatic marking of recitation proficiency the first time in this paper.Based on characteristic and rule analysis for recitation exams,a ＂parallelized sentence network combining silence and filler loop networks＂ was proposed for automatically recognizing the speech text of recitation.The marking features better suiting the recitation exams were extracted from the recognition result,and then by associating them with adaptive techniques in reading proficiency marking,we constructed the first automatic marking system for recitation proficiency.Experiments on a data set of 342 samples collected from ＂FLTRP＂ Cup Contest of New Concept English Recitation in year 2006 show that the system achieved preferable marking performance：the correlation degree between the human expert marking and computerised marking achieves 0.732 with average marking difference 0.358,which is on 90% level of correlation degree between the experts.On two details of marking indices in completeness and fluency with regard to recitation,the automatic marking ability level even outperforms the expert＇s.

关 键 词：计算机辅助语言学习 背诵题评分 识别网络 

分 类 号：TP311.1[自动化与计算机技术—计算机软件与理论]