QQ客服 官方微博 反馈留言

题名	汉语节律分析及其预测方法的研究
作者	董宏辉
学位类别	工学博士
答辩日期	2006-12-31
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 ; 陶建华
关键词	汉语节律 韵律词 韵律组块 韵律短语 Chinese rhythm Prosodic words Prosodic chunks Prosodic phrases
其他题名	Research on the Analysis and Prediction of Chinese Rhythm
学位专业	模式识别与智能系统
中文摘要	本文在对汉语节律规律进行分析的基础上，制定出由节律的稳定成分（韵律词和韵律组块）确定节律结构的方案；确定从语法词预测韵律词的模型方案，并用最大熵方法建立韵律词的预测模型；尝试采用多种机器学习方法构建韵律组块的自动分析器，最终确定采用条件随机场对韵律组块进行建模；为了得到有效的韵律组块分析器，我们还定义了基于文本的韵律组块标注规范，并建立了大规模的韵律组块库；本文还采用信源信道模型作为韵律短语预测的统计模型，利用韵律组块和长度约束建立韵律短语预测的规则模型。具体说来，论文涉及到以下的主要工作： (1) 通过对节律本身的不确定性和规律性的分析，提出了先确定节律中的稳定单元进而实现整个节律预测的框架思想。 (2) 提出了一种基于最大熵模型的韵律词自动预测方法，根据词法信息建立韵律词的预测模型。该模型不仅能够对语法词组合生成韵律词的情况进行建模，而且还能够描述语法词内部分解生成韵律词的情况。使用的特征选择方法，可以选择出有效的特征，大大压缩了模型。 (3) 通过感知试验证明韵律组块在节律组织中的重要性。韵律组块至少可以通过两种途径来影响节律的组织，首先，韵律组块可以进一步减少节律停顿错误的可能；其次，在不破坏韵律组块的前提下，可以组织生成多种停顿方式。试验还证实了节律的长度规划与感知紧密相关。从真实语料文本的韵律组块语料标注入手，制定了两套韵律组块的标注方案。两种标注方案提高了韵律组块标注的效率，为通过统计方法构建韵律组块分析器奠定了基础。 (4) 采用机器学习的方法对韵律组块进行建模，经过试验比较多种统计方法，提出了基于CRF的韵律组块自动识别器。同时，提出了韵律短语识别的方案：基于统计模型的韵律短语识别，基于韵律组块和长度约束的规则模型。并提出多节律模式生成的原型系统，通过切换不同的节律预测系统得到不同节律模式。 以上的研究不仅可以实现汉语节律结构的预测，其基于稳定结构的实现方式可以推广到整个韵律预测领域，具有一定的普遍意义。
英文摘要	Chinese rhythm is one important part for TTS system. This dissertation analyzes the priciples of Chinese rhythm. It is discovered that the prosodic word and the prosodic chunk are the stable units in the rhythm, and the synthesized result based on these stable units also sound natural and fluent. The dissertation includes the following works: (1) Studying the relations between the prosodic structure and the syntactic structure. The freedom and stabilization of the rhythm are also studied. After the analyzation on the rhythm, the framework used to predict the Chinese rhythm is proposed, which is based on the stable units of the rhythm. (2) Proposing a maximum entropy approach to detect the prosodic word in the utterance. This approach can not only model the process of grouping the lexicon words into the prosodic word, but describe how the long lexicon word is split into prosodic words. Using the feature selection technology, the ME model is compressed much small with still high performance. (3) Justifying the importance of the prosodic chunks in the rhythm by a perceptual experiment. The prosodic chunks can influence the rhythm generation through at least two ways: the prosodic chunks can reduce the probability of breaking wrongly; and based on the prosodic chunks, many different acceptable breaking methods for the same sentence can be got by changing the rhythm. The perceptual test also shows that the length of the prosodic phrase will affect the preference score. Two anotaion specs for the prosodic chunk are designed and a large scale corpus of prosodic chunks is built according to these specs. The corpus is the basis for prosodic chunking. (4) Machine learning methods are introduced to model the prosodic chunking. TBL, HMM and CRF are all used for the task. The experiments show that the CRF model is much better than the others. The CRF method is used for the prosodic chunking. The prosodic phrasing is modeled by two ways: a statistic model based prosodic phrasing and a rule based prosodic phrasing. Both methods use the prosodic chunk information. The experiments show that with the prosodic chunk constraints, these two breaking methods are almost equal to each other. Using the prosodic chunk the variable prosodic phrasing system can be done. All the above researches can not only achieve the prosodic structure prediction, but can be used on other prosody problems.
语种	中文
其他标识符	200218014603200
内容类型	学位论文
源URL	[http://ir.ia.ac.cn/handle/173211/5958]
专题	毕业生_博士学位论文
推荐引用方式 GB/T 7714	董宏辉. 汉语节律分析及其预测方法的研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2006.

个性服务

查看访问统计

相关权益政策

暂无数据

收藏/分享

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。