In recent years, thanks to the improvement of computer computing capability and the continuous accumulation of voice data, many new machine learning-based voice processing technologies have emerged,among which, based on the deep neural network algorithm, the end-to-end Tacotron2 voice synthesis system framework has been widely favored in the speech engineering technology. Tacotron2 is an open source program,easy to run, and has been successfully applied to speech synthesis in multiple languages and in different tones.This paper studies the application of Tacotron2 in the Tibetan language and achieves good experimental results. First, a mediumscale database of Tibetan speech（5500 sentences） was constructed through natural speech collection, automatic annotation, and acoustic analysis. These include Tibetan phoneme transformation,digital recording and Mel-spectrum data. Then, the Tibetan speech synthesis test was performed using the open source program Tacotron2 and the above speech database. Last, by error analysis of synthetic Mel-spectrum,and a subjective evaluation of the naturalness of the synthetic speech, the result is shown, that the application of Tacotron2 in the Tibetan language synthesis is effectively reducing the spectral lose, and improving the naturalness of synthetic speech. Therefore, the “end-to-end”-based Tacotron2 synthesis framework has important applications in Tibetan speech synthesis, and deserves further research and promotion.

论文使用mask-predict解码扩展CMLMs模型,研究了藏汉神经机器翻译方法.并且针对该模型进行改进.在不同领域藏汉数据集上,经过实验,与非自回归模型NAT和标准的Transformer模型进行比较,在不降低解码速度的情况下,与非自回归模型相比提升了4个BLEU以上;与标准的Transformer模型相比能够达到甚至超过Transformer模型性能,同时解码速度更快.

文章以目前在前端语言模型中较为突出的字音转换问题为研究对象,结合藏语本身特有的语音特点,一是研究了藏语语音结构,以及语音合成中藏语同音字的发音规律.二是较为全面的统计及分析了藏语同音字的发音归类问题,并根据统计及归类结果,把18738个字中除341个字(没有同音字)之外的其余字归类为2539个同音字类别.三是通过拉丁转写方式对同音字定义了读音标记,给出的2880个音素覆盖了所有18738个藏字.

互联互通时代了解和掌握不同语言的区域文化和信息十分重要,机器翻译是目前广泛应用的交流媒介.本文以藏汉机器翻译为研究对象,利用Transformer框架和模型,研究了基于Transformer多注意力机制的藏汉机器翻译方法.经过实验,评估了多语料融合实验、语料双切分实验对比效果,得到了BLEU值32.6的实验结果.

针对藏文现代图书版面中的文本行分布不均匀、现代藏文字体差异较大的问题,提出了一种基于Faster R-CNN的版面文本行检测算法.通过在整理标注的数据集上训练,用ResNet-50网络提取出藏文现代图书版面特征信息.为了有效提高模型的泛化能力,在COCO数据集下的网络模型中进行迁移学习.实验结果表明,该方法可对藏文现代印刷物的版面实现文本行的定位,检测准确率为83％,召回率为95％,明显提高了版面检测的精确度.

文章以拉萨市纳木措景区为例,通过搜集和分析各大在线旅游平台的游记类评论文本,分析游客对旅游目的地景区的综合形象感知.结果显示:(1)游客对纳木措景区的整体印象出现频率最高的关键词为海拔、雪山、湖水.(2)游客对景区环境、旅行体验、景区服务3个方面的情感得分较高,且多以积极评论为主.(3)游客对景区厕所、停车场、饮食体验方面的评价普遍偏低,且多以消极评论为主.

针对藏文文献版面的特征,本文结合藏字结构的特征,提出一种基于卷积神经网络的DnCNN藏文文献版面去噪算法.该方法采用17层网络结构深度学习版面特征,在训练过程中结合批量标准化和ReLU激活函数提高训练速度和特征的学习精度,最后使用残差学习输出残差图像.实验结果表明,针对藏文文献版面中易出现的噪声有良好的去噪表现,不但有较高的峰值信噪比,而且在主观评价方面也有不错的评价.

自相似特性是复杂网络研究的重点,分形维数是度量其自相似特性的重要工具.针对带权图中节点权重与边权重可以为正实数、负实数、纯虚数及复数等多种不同数值的情形,给出各种不同带权图的多重分形维数,讨论了带权图的多重分形特性.研究表明,在不同类型的带权图中,除节点权重及边权重均为正实数的情形之外,其他类型的带权图均具有多重分形特性.最后分析了这些带权图多重分形维数的性质.

图像在采集或传输的过程中会受到随机噪声的污染,给后续的图像分割、特征提取等处理造成干扰.常见噪声分为椒盐噪声、高斯噪声和混合噪声,为了进行图像处理,需对图像去噪.文章介绍了几种常用的图像去噪方法,如中值滤波、BM3D、均值滤波、小波去噪等,通过对这几种去噪方法的基本思想和算法流程的对比研究,分析并总结了这些方法处理噪声的效果和优缺点.

互联网的高速发展催生了海量网络文本,这对传统的串行文本校对算法提出了新的性能挑战。尽管近年来文本自动校对任务受到了较多关注,但相关研究工作多集中于串行算法,鲜有涉及校对的并行化。文中首先对串行校对算法进行泛化,给出一种串行校对的通用框架,然后针对串行校对算法处理大规模文本存在的耗时长的不足,提出3种通用的文本校对并行化方法:1)基于多线程的线程并行校对,它基于线程池的方式实现段落和校对功能的同时并行;2)基于Spark MapReduce的批处理并行校对,它通过RDD并行计算的方式实现段落的并行校对;3)基于Spark Streaming流式计算框架的流式并行校对,它通过将文本流的实时计算转为一系列小规模的基于时间分片的批处理作业,有效避免了固定开销,显著缩短了校对时延。由于流式计算兼有低时延和高吞吐的优势,文中最后选用流式校对来构建并行校对系统。性能对比实验表明,线程并行适合校对小规模文本,批处理并行适合大规模文本的离线校对,流式并行校对有效减少了约110s的固定时延,相比批处理校对,采用Streaming计算框架的流式校对取得了极大的性能提升。

超网络是较通常意义上的复杂网络更为复杂的网络,该网络的每一条超边能连接任意多个节点的特性使其比复杂网络能更好地描述真实世界中的复杂系统。针对现有超网络研究中对超网络度量方法的缺陷与不足,提出了一种超网络度量方法——超网络维数（HD）,即为所有超边包含的节点权重之和与对应超边权重乘积和的对数值和节点权重之和与超边权重之和乘积对数值的比值的两倍。超网络维数可以应用于节点权重与超边权重为正实数、负实数、纯虚数,乃至复数等多种不同数值类型的带权超网络中。最后给出了超网络维数的若干性质。

随着深度学习的快速发展,目标检测技术已经越来越成熟.目前神经网络的目标检测技术已经被广泛应用在行人检测、自动驾驶、视频追踪等领域.本文首先分析了目标检测技术的发展背景,介绍了神经网络的发展历程和目标检测中常用的卷积神经网络模型;其次分析了深度神经网络下的目标检测方法,指出现阶段不同检测方法存在的问题以及今后我们需要改进的方向;最后对深度神经网络的目标检测技术进行了总结和展望.

文章针对学生时间碎片化的学习特点,分析了翻转课堂教学模式下的微课设计概念、微课课堂教学的内容,提出了坐标式的微课课堂教学模式,探讨了微课设计方法和教学组织方法.

Correlation matrix describes hypernetwork briefly and intuitively.Hypemetwork can be characterized by node degree,node hyperdegree and hyperedge degree.This paper studies hypemetwork especially self-similar hypernetwork and random hypemetwork from the perspective of correlation matrix,and shows several properties of approaches for constructing hypemetwork based on matrix operation.Self-similar hypemetwork can be obtained by Tracy-Singh product on the correlation matrix of a simple initial hypergraph iteratively,and random hypemetwork can be obtained by Tracy-Singh sum on the correlation matrixes of multiple simple initial hypergraphs sequentially.The fractal dimension of self-similar hypernetworks is no larger than 2.When the initial hypergraph is a connected and nonbipartite hypergraph,the diameter of self-similar hypernetwork does not exceed twice of that of the initial hypergraph,namely,it also shares a small-world property.The distributions of node degrees,node hyperdegrees and hyperedge degrees of random hypemetworks are normal.The results of simulation experiments validate the properties of the constmcted hypemetwork.

Part of Speech (POS) tagging is fundamental to Tibetan processing,with a wide applications in Tibetan text classification,information retrieval,machine translation and other fields.This paper proposes a method of Tibetan POS tagging based on distributed representation.First,this method extends the dictionary by semantic approximation according to the distributed representation.Then the POS tagging is completed according to the dictionary and the semantic similarity.Experimental results show that this method can expand the dictionary with a better result.

The method of growth and preferential attachment applied by the classic BA scale-free network model to deal with connections between nodes of network will result in unlimited connections and other defects.This paper improves the method of connections of BA network model by introducing the maximum number of connections, having a sub-linear growth in the number of connections of new nodes and using Logistic function.Then a new network model named BE with a bimodal degree distribution is obtained.Its several properties are also provided.This model may be applied to explain the socio-economic polarization in the real world well.Moreover, the shifting and zooming of the peak may be achieved by adjusting its parameters.BE network model will be degenerated to BA network model in the limiting case.

For the lack of corpus issue in personal relation,this paper studied the methods of automatic tagging based on HUDONG pedia;for poor ability to express feature issue in shallow machine learning models,we proposed the method of personal relation extraction under deep learning model SDAs and focused on the effect of personal relation extraction with combination features and effect of personal relation extraction with different depths in SDAs network.F factor can reach 73.75％ through experiment analysis.

Named entity recognition is an important research content in text mining.It has a high recognition rate by use of statistical principle.This paper studies Tibetan name recognition technology using conditional random fields (CRF)principle,focuses on analysis of the internal structure of the Tibetan names,contextual features,feature selection and data preprocessing,etc.and evaluates the effectiveness of different features through experiments.The contributions of this paper are that the method of name recognition based on the information of word(syllable)and word position is firstly presented;trigger words,function words,dictionary of names and personal noun suffix as features,together with their different combinations and optimization are studied,and the role of the different function words to the name recognition is refined.Experimental evaluation on different combinations showed that:1 )the features of trigger words and ergative particle can play a positive role on the Tibetan name recognition;2)different feature window sizes have an impact on the name recognition;3)the recognition rate of Tibetan names can reach 80%of F 1 value by use of CRF.However,it can’t reach similar recognition results in other languages due to the high am-biguity of words consisting of two Tibetan syllables.

作为文献计量学重要定律的Zipf定律已在许多领域得到较广泛的应用,网络信息计量学伴随着网络信息的激增而受到人们越来越大的关注。该文结合搜索结果数量的分布情况,提出了在网络信息计量学中仍然存在Zipf定律的猜想,并采用公开的词语集在几个代表性的搜索引擎中进行实验验证,证实了搜索结果数目近似服从Zipf定律的结论,其中Baidu与So搜索结果的Zipf指数为0.003。

为了降低原始文本特征空间的维数，获得较高的分类精度与执行效率，对多种文本特征提取方法进行了研究，如卡方、互信息、信息增益、主成分分析（PCA）等。针对传统文本特征抽取方法存在的精度不高、执行效率低等问题，提出了一种基于分块主成分分析的文本特征提取算法。该算法通过 K-均值词聚类进行特征词分块，再对各分块实施 PCA 操作抽取出更具代表性的特征项，最后使用支持向量机分类器对文本进行分类。实验结果表明：分块主成分分析的分类指标 Fβ＝1达到了88．7％，执行时间为353 s，能够有效提高文本分类精度与执行效率。