CORC  > 自动化研究所  > 中国科学院自动化研究所  > 毕业生  > 博士学位论文
题名口语测试自动评估技术研究
作者江杰
学位类别工学博士
答辩日期2009-06-05
授予单位中国科学院研究生院
授予地点中国科学院自动化研究所
导师徐波
关键词口语自动评估 二次多模型强制切分 增强后验概率向量 文本相似度计算 向量空间模型 automatic language proficiency evaluation two time force alignment with multi-models enhanced posterior probability vector text similarity calculation vector space model
其他题名Research on Automatic Scoring for Language Proficiency Evaluation
学位专业模式识别与智能系统
中文摘要口语测试自动评估技术是计算机辅助语言学习领域的重要研究内容。本文针对该领域中朗读与问答题型的自动评估技术做了深入的研究,其中包括朗读和问答题型的自动评分和发音错误的自动诊断。 当前主流的朗读题目的自动口语评估,主要采用的是基于准确率、发音、声调、语速、重音等多特征融合的方法;而主流的发音错误自动诊断技术采用的是声学后验概率计算的GOP方法;此外,目前还没有专门的研究涉及到问答题型的自动评估。在此基础上,本文在该领域的主要贡献和创新点归纳如下: 1. 提出了二次多模型强制切分计算后验概率的方法,用于发音错误自动诊断。通常的GOP方法对于音素边界比较模糊的发音难以得到较好错误检测结果,而本文提出的多模型二次强制切分后验概率计算法,能够明显提高发音错误自动检测系统的性能。 2. 提出了混淆矩阵增强的后验概率向量方法,用于发音错误自动诊断。对于基于后验概率的方法诊断性能较差的音素集合,本文利用混淆矩阵增强的后验概率向量对音素建模,从而获得对应的音素分类器,该音素分类器能够有效地提高发音错误自动诊断的性能。 3. 提出了基于文本相似度计算的语义评估方法,用于问答题型的自动语义评估。该方法通过对问答题上下文文本建立有限状态机对齐模型,之后在对齐数据上计算基于词语相似度的特征作为自动语义评估得分。该方法在标注文本的测试集合上能够达到和人工评分相近的评估结果。 4. 提出了基于向量空间模型的N元语法语义评估方法,用于问答题型的自动语义评估。该方法将测试语音通过识别得出的音素混淆网络映射到向量空间模型中,并提取其与参考答案之间的多维相似度特征,最后拟合得出语义评估得分。该方法在实验中可以达到较好的语义评估效果。 5. 建立了朗读和问答题的评估系统,其中同时包含了发音错误自动诊断功能,并将其应用于实际的口语测试评估项目,从实践的角度有力地验证了本文方法的可行性与有效性,并取得了比较满意的结果。
英文摘要Automatic language proficiency evaluation is one of the vital tasks in Computer Assisted Language Learning. This paper is concerned with the research of automatic scoring for reading and QA questions in automatic language proficiency evaluation. Details are examined on three research topics: automatic scoring for reading questions, automatic mispronunciation diagnosis and automatic scoring for QA questions. Proposed by the state of art technologies, features, such as pronunciations, tones, rate of speeches and stress, are fused to score reading questions, and acoustic posterior probabilities, such as GOP, are utilized in automatic mispronunciation diagnosis. Besides, few researches have addressed the automatic scoring for QA questions. On these bases, the contribution and novelty of this paper are listed as: 1. Proposed the two time force alignment with multi-models for automatic mispronunciation diagnosis. The detection performance is clearly outperformed the traditional GOP based methods. 2. Proposed the enhanced posterior probability vectors (EPPV) with confusion matrix for automatic mispronunciation diagnosis. Multi-classifiers for posterior probability patterns are built to assure the detection performance on the phone sets which are barely indistinguishable for traditional posterior probability based methods. 3. Proposed text similarity based semantic scoring for QA questions. From the context QA questions, finite state machines (FSM) are utilized to build the alignment models and semantic features are extracted by word similarity calculation. The scoring accuracy is comparable with human raters on testing corpus. 4. Proposed N-gram based vector space model (VSM) semantic scoring for QA questions. Phone confusion networks are generated to extract the semantic features in N-gram based vector space models and the fused scores are validated in experiments. 5. Built a system to carry on the automatic scoring for reading and QA questions. Application on real language proficiency evaluation validates its feasibility and effectiveness.
语种中文
其他标识符200618014628061
内容类型学位论文
源URL[http://ir.ia.ac.cn/handle/173211/6219]  
专题毕业生_博士学位论文
推荐引用方式
GB/T 7714
江杰. 口语测试自动评估技术研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2009.
个性服务
查看访问统计
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。


©版权所有 ©2017 CSpace - Powered by CSpace