A feature selection-based speaker clustering method for paralinguistic tasks

In recent years, computational paralinguistics has emerged as a new topic within speech technology. It concerns extracting non-linguistic information from speech (such as emotions, the level of conflict, whether the speaker is drunk). It was shown recently that many methods applied here can be assisted by speaker clustering; for example, the features extracted from the utterances could be normalized speaker-wise instead of using a global method. In this paper, we propose a speaker clustering algorithm based on standard clustering approaches like K-means and feature selection. By applying this speaker clustering technique in two paralinguistic tasks, we were able to significantly improve the accuracy scores of several machine learning methods, and we also obtained an insight into what features could be efficiently used to separate the different speakers.     

An ontology‐based framework for automatic topic detection in multilingual environments

The detection of topics from large textual data volumes is currently a research area, which has many applications in the development of computational systems. A proposed solution for the detection of topics in data mining is the application of clustering methods. This paper presents the application of a new ontology‐based methodology for the automatic topic detection without any previous information based on the use of hierarchical clustering algorithms and a multilingual knowledge base. The approach also includes lexical resources that allow us to enrich the semantics of the analyzed texts. The novelty of this approach consists of the dimensionality reduction of the terms present in the texts by using ontology and the introduction of a method for the creation of a term weight matrix for use in clustering algorithms. With this approach, it is possible to improve automatic topic detection in documents. The proposed methodology was assessed with four datasets (two of them in English and two in Spanish).     

汉语形容词的自动词义区分研究

词义知识获取是词义知识库建设、词义消歧等任务的基础和起点,目前该工作基本依赖人类专家的智慧和洞察力,在大规模文本处理上缺乏意义计算的客观性和一致性。该文以汉语的中高频形容词为样本,深入挖掘词义特征并采用有参数初始化过程的EM迭代算法,实现了从真实文本中自动发现并区分词语词义的过程。该词义区分算法选取易获取的词形特征、基于大规模语料的搭配特征、基于网络语料的属性—宿主关系特征,替代以往难以获取的句法结构特征,并进一步利用HowNet优化了词形特征的选择。该工作可以应用于信息检索等领域,能够对现有词典起到修改和补充的作用,该思路亦可扩展到其他汉语词类上去。     

文本主题的自动提取方法研究与实现

在深入分析了当前流行的文本主题提取技术和方法的基础上,将语义方法融入统计算法,提出了一种基于统计的主题提取方法,并描述了它的实现过程。该方法利用文档内句子之间的语义相关性,实现了文本主题的自动生成。首先对文本进行切词和分句处理实现信息分割,再结合文本聚类技术对文本句进行聚类实现信息合并,最后从每类中抽取代表句生成文本主题。实验结果表明,该方法是一个有效、实用的方法。     

基于上下文三音素DBN模型的连续语音识别

考虑连续语音中的协同发音问题,提出基于词内扩展的单流上下文相关三音素动态贝叶斯网络（SS-DBN-TRI）模型和词间扩展的单流上下文相关三音素DBN（SS-DBN-TRI-CON）模型。SS-DBN-TRI模型是Bilmes提出单流DBN（SS-DBN）模型的改进,采用词内上下文相关三音素节点替代单音素节点,每个词由它的对应三音素单元构成,而三音素单元和观测向量相联系;SS-DBN-TRI-CON模型基于SS-DBN模型,通过增加当前音素的前音素节点和后音素节点,构成一个新的词间扩展的三音素变量节点,新的三音素节点和观测向量相联系,采用高斯混合模型来描述,采用数字连续语音数据库的实验结果表明：SS-DBN-TRI-CON具备最好的语音识别性能。     

一种新的基于主题的语言模型自适应方法

基于主题的语言模型自适应方法应尽可能提高语言模型权重系数的更新速度并降低语言模型的调用量以满足语音识别实时性要求。本文采用基于聚类的方法实现连续相邻二元词对的量化表示并以此刻画语音识别预测历史和各个文本主题中心,依据语音识别历史矢量和各个文本主题中心矢量的相似度更新语言模型权重系数并摒弃全局语言模型。同传统的基于EM算法的自适应方法相比,实验表明该方法明显提高了语音识别性能和实时性,识别错误率相对下降5.1% ,说明该方法可比较准确地判断测试内容所属文本主题。     

基于双层网格和密度的数据流聚类算法

传统的基于网格的数据流聚类算法在同一粒度的网格上进行聚类,虽然提高了处理速度,但聚类准确性较低。针对此问题,提出一种新的基于双层网格和密度的数据流聚类算法DBG Stream。在2种粒度的网格上对数据流进行聚类,并借鉴CluStream算法的思想,将聚类过程分为2个阶段。在线过程中利用粗粒度的网格单元形成初始聚类,离线过程中在细粒度网格单元上,对位于簇边界的网格单元进行二次聚类以提高聚类精度,并实现了关键参数的自动设置,通过删格策略提高算法效率。实验结果表明,DBG Stream算法的聚类精确度较D Stream算法有较大提高,有效解决了传统基于网格聚类算法的聚类精度较低的问题。     

基于自动编码器的中文词汇特征无监督学习

大规模未标注语料中蕴含了丰富的词汇信息,有助于提高中文分词词性标注模型效果。该文从未标注语料中抽取词汇的分布信息,表示为高维向量,进一步使用自动编码器神经网络,无监督地学习对高维向量的编码算法,最终得到可直接用于分词词性标注模型的低维特征表示。在宾州中文树库5.0数据集上的实验表明,所得到的词汇特征对分词词性标注模型效果有较大帮助,在词性标注上优于主成分分析与k均值聚类结合的无监督特征学习方法。     

Classifier-based non-linear projection for adaptive endpointing of continuous speech

In this paper we present an algorithm for segmenting or locating the endpoints of speech in a continuous signal stream. The proposed algorithm is based on non-linear likelihood-based projections derived from a Bayesian classifier. It utilizes class distributions in a speech/non-speech classifier to project the signal into a 2-dimensional space where, in the ideal case, optimal classification can be performed with a simple linear discriminant. The projection results in the transformation of diffuse, nebulous classes in high-dimensional space into compact clusters in the low-dimensional space that can be easily separated by simple clustering mechanisms. In this space, decision boundaries for optimal classification can be more easily identified using simple clustering criteria. The segmentation algorithm proposed utilizes this property to determine and update optimal classification thresholds continuously for the signal being segmented. The performance of the proposed algorithm has been evaluated on data recorded under extremely diverse environmental noise conditions. The experiments show that the algorithm performs comparably to manual segmentations even under these diverse conditions.     

基于Dialogic语音卡实时数据采集的电话语音识别系统

语音识别技术在新一代呼叫中心的IVR系统中得到了广泛的应用。为了使用Dialogic电话语音卡进行语音识别,文章解决了用Dialogic语音卡进行语音数据实时采集的问题,并给出了一种用动态背景噪声电平检测语音的算法,建立起了基于DialogicD/120JCT-LS电话语音卡的自动电话交换转接系统。     

基于深度神经网络的关键词识别系统

针对当前关键词识别少资源或零资源场景下的要求, 提出一种基于音频自动分割技术和深度神经网络的关键词识别算法. 首先采用一种基于度量距离的改进型语音分割算法, 将连续语音流分割成孤立音节, 再将音节细分成和音素状态联系的短时音频片段, 分割后的音频片段具有段间特征差异大, 段内特征方差小的特点. 接着利用一种改进的矢量量化方法对音频片段的状态特征进行编码, 实现了关键词集内词的高精度量化编码和集外词的低精度量化编码. 最后以音节为识别单位, 采用压缩的状态转移矩阵作为音节的整体特征, 送入深度神经网络进行语音识别. 仿真结果表明, 该算法能从自然语音流中较为准确地识别出多个特定关键词, 算法易于理解、训练简便, 且具有较好的鲁棒性.     

基于维度最大熵数据流聚类的异常检测方法

针对传统数据流聚类算法聚类信息损失大、不准确的缺点, 提出一种基于维度最大熵的数据流聚类算法. 采用动态数据直方图将数据维度划分为不同的维度组, 计算各维度最大熵划分维度空间簇, 将相同维度簇的数据聚集成微簇, 通过比较微簇的信息熵大小及其分布特点实现数据流的异常检测. 该方法提升了聚类速度, 克服了传统数据流聚类算法信息丢失的缺点. 实验结果表明, 所提出算法能够提高数据流异常检测的准确性和有效性.

     

基于主题的Web文本聚类方法

针对传统Web文本聚类算法没有考虑Web文本主题信息导致对多主题Web文本聚类结果准确率不高的问题,提出基于主题的Web文本聚类方法。该方法通过主题提取、特征抽取、文本聚类三个步骤实现对多主题Web文本的聚类。相对于传统的Web文本聚类算法,所提方法充分考虑了Web文本的主题信息。实验结果表明,对多主题Web文本聚类,所提方法的准确率比基于K-means的文本聚类方法和基于《知网》的文本聚类方法要好。     

Unsupervised Pattern Discovery in Speech

We present a novel approach to speech processing based on the principle of pattern discovery. Our work represents a departure from traditional models of speech recognition, where the end goal is to classify speech into categories defined by a prespecified inventory of lexical units (i.e., phones or words). Instead, we attempt to discover such an inventory in an unsupervised manner by exploiting the structure of repeating patterns within the speech signal. We show how pattern discovery can be used to automatically acquire lexical entities directly from an untranscribed audio stream. Our approach to unsupervised word acquisition utilizes a segmental variant of a widely used dynamic programming technique, which allows us to find matching acoustic patterns between spoken utterances. By aggregating information about these matching patterns across audio streams, we demonstrate how to group similar acoustic sequences together to form clusters corresponding to lexical entities such as words and short multiword phrases. On a corpus of academic lecture material, we demonstrate that clusters found using this technique exhibit high purity and that many of the corresponding lexical identities are relevant to the underlying audio stream.     

一种基于聚类的汉语词语知识的获取方法

在自然语言处理中,知识的自动获取一直是一个核心问题。但如何实现知识的自动获取呢?该文在基于实例的学习方法(Example-BasedLearning,简称EBL)的基础上,提出了一种基于聚类的汉语多义词知识的自动获取方法。实验结果证明,用该方法获得的知识对汉语的词义排歧是有效的。     

基于句子级的唇读语料库及其切分算法

论文对适合唇读研究的连续音节双模态语料库及其语料切分算法的设计和研究工作进行了讨论。介绍了基于句子级的双模态语料库HITBi-CAVDatabaseII的设计和建立,形式化地讨论了该库的主要特点及基于语音能量的语料切分算法的可行性。该切分算法在基于能量的语音切分算法基础上,结合了双模态语料库的一些特征,实现了对语料的自动切分。     

融合距离度量和高斯混合模型的中文词义归纳模型

词义归纳是解决词义知识获取的重要研究课题,利用聚类算法对词义进行归纳分析是目前最广泛采用的方法。通过比较K-Means聚类算法和EM聚类算法在 各自 词义归纳模型上的优势,提出一种新的融合距离度量和高斯混合模型的聚类算法,以期利用两种聚类算法分别在距离度量和数据分布计算上的优势,挖掘数据的几何特性和正态分布信息在词义聚类分析中的作用,从而提高词义归纳模型的性能。实验结果表明,所提混合聚类算法对于改进词义归纳模型的性能是十分有效的。     

基于语音流畅度与模糊聚类的精神分裂症自动识别

针对精神分裂症诊断周期长、缺乏客观诊断依据的问题,提出语音流畅度矩形参数过能熵积,结合模糊聚类双重定位停顿区域的算法辅助诊断精神分裂症.该算法综合了精神分裂症患者语音流畅度低及能量平缓的特征,定位精神分裂症语音的停顿区域,提出语音流畅度量化参数提取算法,结合SVM(支持向量机)分类器,实现精神分裂症的自动识别.提取28例精神分裂症患者和28例正常对照组语音的流畅度声学特征,自动识别精神分裂症,正确率为85％以上.提出的基于过能熵积与模糊聚类的精神分裂症自动识别算法,能为临床诊断精神分裂症提供客观、有效、无创的辅助依据.     

TOPIC DETECTION OF UNRESTRICTED TEXTS: APPROACHES AND EVALUATIONS

ABSTRACT

Topic detection and tracking refers to automatic techniques for locating topically related cohesive paragraphs in a stream of text. Most documents are about more than one subject, but many Natural Language Processing (NLP) and Information Retrieval (IR) techniques implicitly assume documents have just one topic. Even in the presence of a single topic within a document, the document may address multiple subtopics and various aspects of the primary topic. Hence, dividing documents into topically coherent units and discovering their topic might have many uses. We describe new clues that account for the topic of grouping of contiguous portions of the text. Those clues are based on general lexical resources, which make them applicable to unrestricted texts, and can have many uses such as helping users find answers to general questions in an information search task, or in question/answering systems, or in text summarization. We devise an algorithm for identifying these clues, and we report on the performance of these clues, as well as the improvements suggested by our experiments.