基于CART技术的汉语韵律短语分析

韵律短语分析是文语转换(Text-to-Speech,TTS)系统中的一项重要内容。它负责在输入文本的不同位置上,标记上不同级别的符号形式的韵律短语边界,从而把一个句子分隔成不同层次的短语。韵律短语分析的结果直接影响着TTS系统的后继步骤,如韵律参数的生成。杨玉芳研究了不同级别的韵律短语边界与附近音节的时长、停顿、基频等韵律参数之间的关系,考察了这些参数随边界等级的系统变化。此外,在口语理解等应用系统中,韵律短语分析也有可用之处。因为口语理解的一个难点在于自然语言存在着歧义现象:即便是完全相同的一个句子,如果在表述时断句的位置不同,就可能产生完全不同的含义,所以有的研究者就利用句子的韵律边界信     

基于约束模型的韵律短语预测

本文提出了基于语法约束和长度约束的韵律短语预测模型。在语法约束模型中,我们引入了组块作为基本的节律分析单元。韵律短语的长度约束模型是利用隐马尔科夫模型对语句中韵律短语的长度规划进行建模,这个模型对短语的长度分布及韵律词与韵律短语的关系进行了描述。最后,利用一个称为k-候选的方法来融合这两个约束模型。整个方法充分利用了韵律短语的语法约束和长度约束,并将之有机地结合起来。试验表明,该预测模型达到了很好的效果,韵律短语边界识别的调和平均值达到82.9%。     

汉语韵律短语切分研究

首先分析了汉语中韵律短语的边界情况，然后提出了一个韵律短语切分方法，开放和封闭测试表明该方法取得了相当好的结果。     

中文语音合成系统中的一种两层韵律结构生成体系

韵律结构生成是改进一个语音合成系统中的合成语音的完整度和自然度的重要组成部分. 韵律词和韵律短语的自动切分是中文层级韵律结构的两个重要的基本层面, 本文调研了这个基本问题, 并提出了一种两层韵律结构生成体系. 为此, 我们建立了条件随机场模型为韵律词和韵律短语的预测选取不同的前端特征. 除此之外, 我们还引入了基于转换的错误驱动学习模块来修正后端的初始预测. 实验结果显示, 这种结合条件随机场和错误驱动学习的方法使得韵律词和韵律短语的自动分割的F-score值达到了94.66%.     

基于韵律特征和语法信息的韵律边界检测模型

韵律短语边界的自动检测,对语音合成中语料库的韵律标注以及语音识别中韵律短语的自动划分都有重要意义。本文通过对影响韵律短语边界的声学、韵律等参量的分析,得到和韵律短语边界关联性较大的一组声学特征参数、韵律环境参数和语法信息;同时引入语音合成中的韵律预测思想,在假定所有音节边界均为非韵律短语边界时,预测每个音节的基频。最后使用决策树模型,将音节边界处的韵律环境信息、语法信息以及预测结果作为决策树的输入,利用决策树综合判定当前音节边界是否为韵律短语的边界。实验表明,这种方法对于基于确定性文本(text-dependent)的语音韵律短语边界的检测,具有较好效果,同时可以显著提高语音合成中语料库的标注效率和标注结果的一致性。     

基于边界点词性特征统计的韵律短语切分

由于基于规则方法的文本处理系统在系统建立时需要总结大量的规则,而且很难保证它在处理大规模真实文本时的强壮性,因此本文在使用统计方法进行韵律短语切分方面做了一些有益的探索。先对文本进行自动分词和自动词性标注,然后利用从已经经过人工标注的语料库中得到的韵律短语切分点的边界模式以及概率信息,对文本中的韵律短语切分点进行自动预测,最后利用规则进行适当的纠错。通过对一千句的真实文本进行封闭和开放测试, 词性标注的正确率在95%左右,韵律短语切分的召回率在60%左右,正确率达到了80%。     

基于语块的英语文语转换系统的韵律生成方法

为预测英语文语转换(Text-to-Speech,TIS)系统中韵律生成模块的韵律边界,通过在中间短语、语调短语和语句后分别插入不同长度的停顿,产生使合成语音具有与真人语音类似的韵律结构.通过采用基于语块的中间短语切分,以中间短语为基本单位,生成一个语调短语边界预测的学习语料库,然后采用转换式学习法进行标注学习,从而实现韵律边界的切分.在对真人语料库进行测试的实验中,标注正确率达到81.32%,通过在学习中增加语调短语音节数和标点符号的约束规则,可进一步提高标注正确率.     

汉语韵律短语的时长与音高研究

语句和篇章的韵律结构和信息结构的分析及模型化是提高语音合成的自然度、降低自然语言识别错误率的关键。该文在带有韵律标注ASCCD语料库的基础上对韵律短语的时长和音高特性进行了研究,得到并验证了如下一些结论:(1)韵律短语边界对音节时长有明显的延长作用,不同声调对音节的时长延长作用不同,并且不同的重音级别对音节时长的延长作用也不同。(2)韵律短语边界处中断的时长在较小的韵律边界表现的更为明显。韵律短语的边界处发生了明显的音高重置现象,韵律短语的音高低线总是下降的,而音高高线只是在重音后下降,并且重音处的音域大而且音高高线的位置高。     

一种改进的基于决策树的英文韵律短语边界预测方法

在英文语音合成系统中,韵律短语边界预测的精度对合成语音的自然度和可懂度有着至关重要的影响。基于决策树的预测方法是现阶段最为常用的韵律短语边界预测方法,但因决策树构建时受到数据平衡性制约,难以针对关键词进行建模,而且在基于决策树进行预测时采用了局部最优的搜索方式无法达到全局最优。所以,为了进一步提升韵律短语边界的预测效果,对基于决策树的预测方法进行了改进,引入韵律短语条件概率,使用Viterbi算法同时优化韵律短语边界概率和条件概率,并提出了基于关键词在韵律短语中的位置分布特性的决策树节点概率优化方法。实验表明,在基线系统上使用改进方法后,F-Score由68.7%提升到77.8%,而不可接受率从22.4%降低到15.2%。     

维吾尔语情感语音韵律转换研究

面向维吾尔语情感语音转换,提出一种韵律建模转换方法。该方法结合了维吾尔语韵律特点及语言特点,首次利用离散余弦变换（DCT）分别参数化维吾尔语音节和韵律短语的情感基频。采用高斯混合模型（GMM）训练中性-情感基频联合特征,同时合成中性语速情感语音和情感语速情感语音,主观评测结果显示情感语速更有助于表达情感效果。主客观实验结果显示转换方法可有效进行维吾尔语情感韵律转换,三种情感下,音节和韵律短语的结果均达到75%以上,韵律短语的转换效果要稍优于音节。     

K(Knowledge)-net: building up and its dynamics

The essence of intelligence is to use certain abilities to obtain knowledge, to use that knowledge, and to operate with that knowledge. New knowledge learned by a human is often related to old existing knowledge, and sometimes we could have more conceptual knowledge based on old knowledge. So, the knowledge in the brain exists in a related structural form, and this structure is dynamic, and therefore is evolvable. Based on the understanding of the real process of learning by a human being, we discuss how to make a model to describe the dynamic structure of knowledge. This model is also a principle of artificial brain design. Most of the knowledge a child learns is from natural language and perception information, and we define this as semantic knowledge. The model to describe the process and structure of knowledge growing in a network form is called a K-net. It is a dynamic network with two main dynamics: one is new knowledge added, and the other is aggregating knowledge existing in the network with some probability. Under these very natural conditions, we found that the network is originally a simple random net, and then some characteristics of a complex network gradually appear when more new knowledge is added and aggregated. A more interesting phenomenon is the appearance of a random hierarchical structure. Does this mean emergence?     

基于通用比较模型的t-可诊断系统的特征化及并行诊断算法

文中介绍了一种系统级故障诊断模型－－通用比较模型,该模型处理器作为自身的比较器,综合经典的ＰＭＣ模型和Ｍａｅｎｇ／Ｍａｌｅｋ模型的优点,基于该报多处理器系统的ｔ＿可诊断性问题,给出了ｔ＿可诊断系统的特征化,证明了一个系统成为ｔ－可诊断系统的新的充分必要条件,其次,证明在通用比较模型中,确定故障顺集的问题等价于求解一个超图的最小横切集（Ｍｉｎｉｍｕｍ ｔｒａｖｅｒｓａｌ）,该超图是根据多处理器的通信     

An Order-Based Theory of Updates for Closed Database Views

The fundamental problem in the design of update strategies for views of database schemata is that of selecting how the view update is to be reflected back to the base schema. This work presents a solution to this problem, based upon the dual philosophies of closed update strategies and order-based database mappings. A closed update strategy is one in which the entire set of updates exhibit natural closure properties, including transitivity and reversibility. The order-based paradigm is a natural one; most database formalisms endow the database states with a natural order structure, under which update by insertion is an increasing operation, and update by deletion is decreasing. Upon augmenting the original constant-complement strategy of Bancilhon and Spyratos – which is an early version of a closed update strategy – with compatible order-based notions, the reflection to the base schema of any update to the view schema which is an insertion, a deletion, or a modification which is realizable as a sequence of insertions and deletions is shown to be unique and independent of the choice of complement. In addition to this uniqueness characterization, the paper also develops a theory which identifies conditions under which a natural, maximal, update strategy exists for a view. This theory is then applied to a ubiquitous example – single-relational schemata constrained by equality-generating dependencies. Within this framework it is shown that for a view defined as a projection of the main relation, the only possibility is that the complement defining the update process is also a projection, and that the reconstruction is based upon functional dependencies.     

基于Struts和JDO的信息管理系统

Struts是广泛应用的Web框架，但它没有提供如何访问数据的方法。JDO是数据访问框架，利用O-R映射来访问数据库，而一般要用JDBC和SQL来访问数据库，较之JDO更加方便和快捷。一个成熟的应用需要两者结合起来才能形成一个完整的方案。这里提出用Struts作为前台和逻辑控制，用JDO作为后台数据访问层来开发B/S架构的信息管理系统的一种方案。     

二进神经网络中汉明球的逻辑意义及一般判别方法

剖析二进神经元的逻辑意义对二进神经网络的规则提取是十分重要的，在布尔空间中，汉明球是一种线性可分的空间结构，如何从汉明球中提取出具有清晰逻辑意义的规则是二进神经网络研究的一个问题，通过对MofN规则表达形式的扩展，分析了汉明球的逻辑意义，提出了表达汉明球逻辑意义的LEM规则和GEM规则方法，并且讨论了汉明球和汉明补球的等价性，另一个重要的结果是证明了二进神经元和汉明球等价的充要条件，从而建立了判别汉明球的一般方法。     

混合关系中的函数依赖及其相关规则

在实际应用中有这样一类关系数据库,其中数据项在某个属性上的取值本身又是一个关系。讨论这种混合关系中的函数依赖及其相应的相关规则具有一定的理论意义和实际应用价值。给出这种混合关系的形式定义、混合关系中4种类型的函数依赖关系以及相应的4种类型的相关规则;并给出混合关系到一般关系的转换,通过这个转换说明了混合关系与普通关系的联系及其差别。给出一个实际应用中的例子,来说明混合关系中的函数依赖和相关规则的形式多样性。     

Correlation between learning (probability of success) and fuzzy entropy in control of intelligent robot's part macro-assembly tasks with sensor fusion techniques

A correlation between a learning and a fuzzy entropy, using the control of robotic part macro-assembly (part-bringing) task as an example, is introduced. Two intelligent part-bringing algorithms, to bring a part from an initial position to an assembly hole or a receptacle (target or destination) for a purpose of a part mating in a partially unknown environment containing obstacles, related to a robotic part assembly task are introduced. An entropy function, which is a useful measure of the variability and the information in terms of uncertainty, is introduced to measure its overall performance of a task execution related to the part-bringing task. The degree of uncertainty associated with the part-bringing task is used as an optimality criterion, e.g. minimum entropy, for a specific task execution. Fuzzy set theory, well-suited to the management of uncertainty, is used to address the uncertainty associated with the macro-assembly procedure. In the first algorithm, a macro-assembly, locating various shaped assembly holes (targets) in the workspace corresponding to the shapes of the parts and then bringing the part to the corresponding target, despite existing obstacles is introduced. This is accomplished by combining a neural network control strategy coordinating with a mobile rectilinear grid composed of optical sensors as well as fuzzy optimal controls. Depending on topological relationships among the part's present position, the position of obstacles, and the target position in the workspace, a specific rulebase from a family of distinct fuzzy rulebases for avoiding obstacles is activated. The higher the probability, the input pattern (or value) of the neural network to be identified as the desired output is, the lower the fuzzy entropy is. Through the fuzzy entropy, a degree of identification between the input pattern and the desired output of the neural network can be measured. In the second algorithm, a macro-assembly with a learning algorithm and a sensor fusion for bringing the part to the target is introduced. By employing a learning approach, the uncertainty associated with the part-bringing task is reduced. The higher the probability of success is, the lower the fuzzy entropy is. The results show clearly the correlation between a probability of success related to the task execution of the part-bringing and the fuzzy entropy, and also show the effectiveness of above methodologies. The proposed technique is not only a useful tool to measure the behaviour of the learning but applicable to a wide range of robotic tasks including motion planning, and pick and place operations with various shaped parts and targets.     

Decentralised control for formation reorganisation of multi-agent systems using a virtual leader

In this article, we study the topology reorganisation of formations for multi-agent systems with a broken agent during motion. The cases with a broken agent in different topologies and different positions are studied, and a decentralised algorithm is proposed to reorganise the different damaged topologies of formations. The main contributions of this article include: first, a collective potential function model is designed to reorganise the topologies, and a repulsive potential function model is proposed to avoid obstacles; second, a decentralised algorithm based on potential functions is presented; third, a virtual agent is designed to be a leader as soon as an agent is damaged, and disappears once the topology is repaired. Finally, some simulation results are studied to demonstrate the efficiency of the proposed algorithm.     

ON THE COMPUTATION OF THE DIGITAL CONVEX HULL AND CIRCULAR HULL OF A DIGITAL REGION

The problems of defining convexity and circularity of a digital region are considered. A new definition of digital convexity, called DL- (digital line) convexity, is proposed. A region is DL-convex if, for any two pixels belonging to it, there exists a digital straight line between them all of whose pixels belong to the region. DL-convexity is shown to be stronger that two other definitions, T- (triangle) convexity and L- (line) convexity. A digital region is T-convex if it is DL-convex, but the converse is not generally true. This is because a DL-convex region must be connected, but T- and L-convex regions can be disconnected. An algorithm to compute the DL-convex hull of a digital region is described. A related problem, the computation of the circular hull and its application to testing the circularity of a digital region, is also considered, and an algorithm is given that is computationally cheaper than a previous algorithm for testing circularity.     

一种基于网格服务路线的异常处理方法

网格服务是提供一组定义明确接口并遵守一定规范的Web服务．网格服务异常(exception)是指在服务请求、响应和执行过程中可能发生的问题．网格服务异常处理与网格服务运行状态、上下文有关．网格服务异常处理的主要内容包括网格服务检查点(checkpoint)设置、服务运行跟踪、以及服务回卷(rollback)．传统处理方法是使用一个错误码：一个函数检测到错误,将错误码传送给调用者．这一处理过程可能要多级连续处理,直到该错误码被传送到响应该错误的处理函数．如果一个函数调用连接比较弱,将会造成整个处理过程失败．网格服务异常处理一方面应尽量减少传统处理单元内部错误代码繁殖的现象,另一方面要考虑服务异常向外部单元或网格服务传送的功能．这里,从网格系统结构出发,首先对网格服务异常处理进行讨论;然后提出了用于网格服务异常处理的A/B—Back算法;最后给出了基于Globus Toolkit 3．0的模拟实验结果分析．