基于特征融合的中文新闻文本情感分类方法研究

针对现有的新闻文本情感分析任务中,单一模型提取文本特征的片面性,且无法充分提取新闻文本语义等特征问题,提出一种基于门控单元特征融合的BERT-CNN情感分析方法。该方法分别采用BERT语言预训练模型与卷积神经网络（CNN）对新闻文本的特征向量进行提取;然后采用门控循环单元对提取到的文本特征进行特征融合;再输入到Softmax层进行新闻文本分类;最后从精准率、召回率和F1-Score三个维度对比BERT、BERT-CNN、BERT-DPCNN和BERT-ERNIE的实验结果。实验结果表明,当分类场景更换为情感识别时,BERT-CNN依旧具有强大的语义捕捉能力,证明了BERT-CNN的泛化能力;另外,从原BERT的提升效果看,基于门控单元特征融合的BERT-CNN方法（提升2.07%）比词向量的方法（提升0.31%）更高。这一结果也证明了基于门控单元特征融合的BERT-CNN方法的有效性。     

基于组合-卷积神经网络的中文新闻文本分类

目前的新闻分类研究以英文居多,而且常用的传统机器学习方法在长文本处理方面,存在局部文本块特征提取不完善的问题.为了解决中文新闻分类缺乏专门术语集的问题,采用构造数据索引的方法,制作了适合中文新闻分类的词汇表,并结合word2vec预训练词向量进行文本特征构建.为了解决特征提取不完善的问题,通过改进经典卷积神经网络模型结构,研究不同的卷积和池化操作对分类结果的影响.为提高新闻文本分类的精确率,本文提出并实现了一种组合-卷积神经网络模型,设计了有效的模型正则化和优化方法.实验结果表明,组合-卷积神经网络模型对中文新闻文本分类的精确率达到93.69％,相比最优的传统机器学习方法和经典卷积神经网络模型精确率分别提升6.34％和1.19％,并在召回率和F值两项指标上均优于对比模型.     

基于特征融合的多方面抽取及情感分析

针对序列标注标签预测空间大导致模型预测效果较差的问题,提出一种基于BERT-BiLSTM-Fusion的多方面抽取及情感分析模型。采用跨度预测的方法进行方面词抽取与方面词情感预测分类并联合训练,通过Bert预训练语言模型得到文本嵌入表示,使用BiLSTM学习观测序列上的依赖关系增强学习位置信息,提高模型抽取效果。对Bert预训练语言模型的每一层输出特征进行特征融合,提高模型的情感极性分类效果,并且在三个公开数据集Laptop、Restaurant、Twitter上设计对比实验以及消融实验。实验结果显示,BERT-BiLSTM-Fusion模型的F1值分别达到了66.72%、78.44%、62.10%,且高于对比模型,表明了所提出模型的有效性。     

基于ERNIE-BiLSTM的社交网络文本情感分析

社交网络文本情感分析任务中,因短文本信息模糊等特点,传统的词向量模型无法更好地表示词的语义特征,当前短文本情感分类任务多以二分类研究为主,将结果分类为积极情感与消极情感,未能对分类结果更细入的划分。文中提出一种舆情情感分析的ERNIE-BiLSTM方法,实现了对用户评论情感的七种情绪分类,包括恐惧、厌恶、乐观、惊喜、感恩、悲伤和愤怒。ERNIE-BiLSTM方法利用ERNIE预训练模型获取文本的语义信息,结合BiLSTM提取文本的双向特征,最后使用softmax函数获得最终的情感分类结果。实验结果表明,ERNIE-BiLSTM方法具有87.7%的精确率、86.9%的召回率和86.8%的F1得分,比其他方法得到了有效提升。     

基于混合排序填充网络的文本到指令序列翻译

本论文主要研究利用语言预训练模型和深度学习,来进行文本序列到指令序列的翻译.在文本序列到指令序列翻译的领域内,优势模型按结构划分主要分为2类,一种是端到端模型,另外一种是在编码器和解码器中间引入中间逻辑表示层.与以往的方法不同,本文提出了名为混合排序填充网络的新模型.该模型将问题与一列单独组合,利用语言预训练模型获取编码,然后根据选择的模板,划分并填充不同的子任务,生成完整的指令序列.这种方法很好地利用了语言预训练模型的特点,避免引入专门的中间语义表示层或者解码器,减少了模型复杂度.在自构建的文本到指令序列数据集上取得了很好的翻译效果,翻译结果的逻辑准确率可达到89.1％.     

基于改进的BERT-CNN模型的新闻文本分类研究

语言处理模型层出不穷,从在图像领域运用的卷积神经网络(Convolutional Neural Networks,CNN)、被改进后的TextCNN到循环神经网络(Rerrent Neural Network,RNN)、谷歌推出的transformer模型以及百度公司提出的ERNIE模型等.为了区分出在新闻文本领域处理效果最佳的文本分类模型,基于新闻文本任务数据对来自变换器的双向编码器表征量(Bidirectional Encoder Representations from Transformers,BERT)进行微调,再将BERT作为embedding输入到其他深度学习模型中,最后对比目前几个深度学习模型的训练效果.实验结果表明,BERT-CNN模型分类效果最佳,其准确率比原BERT模型的准确率多了0.31％,且更为稳定.     

结合双预训练语言模型的中文文本分类模型

针对Word2Vec等模型所表示的词向量存在语义模糊从而导致的特征稀疏问题,提出一种结合自编码和广义自回归预训练语言模型的文本分类方法。首先,分别通过BERT、XLNet对文本进行特征表示,提取一词多义、词语位置及词间联系等语义特征;再分别通过双向长短期记忆网络(BiLSTM)充分提取上下文特征,最后分别使用自注意力机制(Self＿Attention)和层归一化(Layer Normalization)实现语义增强,并将两通道文本向量进行特征融合,获取更接近原文的语义特征,提升文本分类效果。将提出的文本分类模型与多个深度学习模型在3个数据集上进行对比,实验结果表明,相较于基于传统的Word2Vec以及BERT、XLNet词向量表示的文本分类模型,改进模型获得更高的准确率和F₁值,证明了改进模型的分类有效性。     

基于预训练模型的医疗命名实体识别

文中针对生物医学实体识别中存在的边界识别不准确和鲁棒性差的问题,提出了一种融合了预训练语言模型BERT与跨度标签网络的命名实体识别模型。该模型利用BERT获取文本的上下文信息,并结合跨度标签网络进行实体分类及边界判定,显著提升了实体识别的准确性。为增强模型的鲁棒性,引入对抗训练策略,通过迭代训练正常样本与对抗样本,以优化模型参数。基于CCKS2019评测数据集的实验表明,应用对抗训练方法后,其精准率、召回率及F1值均有所提升,验证了对抗训练能对提高模型的预测能力和鲁棒性的有效性。     

基于ABBSAC模型的中文事件抽取方法

事件抽取作为信息抽取的重要一环,是非结构化文本转化为有价值的结构化文本的主要方式。针对目前事件抽取模型普遍训练时间长、模型体量大等问题,提出了一个基于ABBSAC的中文事件抽取模型。通过ALBERT预训练模型缩减模型体量,采用BiSRU++捕捉文本内部关联信息,并融合注意力机制提升模型精度,最后以CRF的输出作为抽取结果。基于新浪新闻自主构建了语料集,进行了对比实验。在获得较高准确率、召回率以及F1值的基础上,该模型训练速度提高了约10%,模型参数量裁剪了约82%,证明了所提模型的先进性。同时,在ACE05和DUEE基准测评数据集上,与前沿方法相比较,将触发词抽取的F1值分别提升了1.7%、0.3%,将论元角色抽取的F1值分别提升了5.4%、0.1%,有效提升了中文事件抽取任务的效能。     

基于BERT-LSTM模型的情感分析研究

情感分类技术在舆论评价和商品评价等诸多领域均有广泛使用,对自然语言处理领域来说具有很重要的研究意义。在当前的社交网络文本中,用户不仅仅使用文字来传达情感,其文本中的表情元素也带有浓厚的情感色彩。传统情感分析模型容易忽略表情元素,从而导致模型未能正确判断文本情感。文中将BERT预训练模型和长短时记忆网络相结合,运用带表情元素的weibo＿senti＿100k数据集实现一个针对微博评论的情感二分类模型。BERT-LSTM模型利用BERT嵌入层对预处理后的句子进行分割并将其转换为动态词向量,结合LSTM模型提取文本和表情元素的特征,最后以预测评论文本的情感极性。实验验证表情元素的重要性和BERT-LSTM模型情感分类的有效性,结果表明同时考虑文字和表情元素相较于纯文字来说模型分类准确率提高20%,BERT-LSTM模型的分类准确率为98.31%、F1值为98.28%,相比传统机器学习模型和其他深度学习模型在最终结果上表现出明显优势。     

Development of the readout system for waveguide multilayer optical card

Waveguide multilayer optical card （WMOC） is a novel storage device of three-dimensional optical information. An advanced readout system fitting for the WMOC is introduced in this paper. The hardware mainly consists of the light source for reading, WMOC, motorized stages addressing unit, microscope imaging unit, CCD detecting unit and PC controlling ＆ processing unit. The movement of the precision motorized stage is controlled by the computer through Visual Basic （VB） language in software. A control panel is also designed to get the layer address and the page address through which the position of the motorized stages can be changed. The WMOC readout system is easy to manage and the readout result is directly displayed on computer monitor.     

Nanoimprint Lithography -A Next Generation High Volume Lithography Technique

IntroductionNanoimprint Lithography is a well-acknowl-edged low cost, high resolution, large area pattern-ing process. It includes the most promising methods,high-pressure hot embossing lithography (HEL) [2],UV-cured imprinting (UV-NIL) [3] and micro contactprinting (m-CP, MCP) [4]. Curing of the imprintedstructures is either done by subsequent UV-lightexposure in the case of UV-NIL or by cooling downbelow the glass transition temperature of the ther-moplastic material in case of HEL…     

Collinear phase-matching study of terahertz-wave generation via difference frequency mixed in GaAs and Inp

The collinearly phase-matching condition of terahertz-wave generation via difference frequency mixed in GaAs and InP is theoretically studied. In collinear phase-matching, the optimum phase-matching wave hands of these two crystals are calculated. The optimum phase-matching wave bands in GaAs and lnP are 0.95-1.38μm and 0.7-0.96μm respectively. The influence of the wavelength choice of the pump wave on the coherent length in THz-wave tuning is also discussed. The influence of the temperature alteration on the phase-matching and the temperature tuning properties in GaAs crystal are calculated and analyzed. It can serve for the following experiments as a theoretical evidence and a reference as well.     

Effects of the alloy compositions on the phonon-polaritons in ternary mixed crystals

Composition dependence of bulk and surface phonon-polaritons in ternary mixed crystals are studied in the framework of the modified random-element-isodisplacement model and the Bom-Huang approximation. The numerical results for Several Ⅱ - Ⅵ and Ⅲ- Ⅴ compound systems are performed, and the polariton frequencies as functions of the compositions for ternary mixed crystals AlxGa1-xAs, GaPxAS1-x, ZnSxSe1-x, GaAsxSb1-x, GaxIn1-xP, and ZnxCd1-xS as examples are given and discussed. The results show that the dependence of the energies of two branches of bulk phonon-polaritons which have phonon-like characteristics, and surface phonon-polaritons on the compositions of ternary mixed crystals are nonlinear and different from those of the corresponding binary systems.     

Influence of PVK ： NPB hole transporting layer on the characteristics of organic light-emitting devices

A doping system consisting of NPB and PVK is employed as a composite hole transporting layer （CHTL）. By adjusting the component ratio of the doping system, a series of devices with different concentration proportion of PVK ： NPB are constracted. The result shows that doping concentration of NPB enhances the competence of hole transporting ability, and modifies the recombination region of charge as well as affects the surface morphology of doped film. Optimum device with a maximum brightness of 7852 cd/m^2 and a power efficiency of 1.75 lm/W has been obtained by choosing a concentration proportion of PVK ： NPB at 1：3.     

New insert layer structure OLEDs

An insert layer structure organic electroluminescent device（OLED） based on a new luminescent material （Zn（salen）） is fabricated. The configuration of the device is ITO/CuPc/NPD/Zn（salen）/Liq/LiF/A1/CuPc/NPD/Zn（salen）/Liq/LiF/A1. Effective insert electrode layers comprising LiF（1nm）/Al（5 nm） are used as a single semitransparent mirror, and bilayer cathode LiF（1 nm）/A1（100 nm） is used as a reflecting mirror. The two mirrors form a Fabry-Perot microcavity and two emissive units. The maximum brightness and luminous efficiency reach 674 cd/m^2 and 2.652 cd/A, respectively, which are 2.1 and 3.7 times higher than the conventional device, respectively. The superior brightness and luminous efficiency over conventional single-unit devices are attributed to microcavity effect.     

Analysis of error performance on Turbo coded FDPIM

Due to variable symbol length of digital pulse interval modulation（DPIM）, it is difficult to analyze the error performances of Turbo coded DPIM. To solve this problem, a fixed-length digital pulse interval modulation（FDPIM） method is provided. The FDPIM modulation structure is introduced. The packet error rates of uncoded FDPIM are analyzed and compared with that of DPIM. Bit error rates of Turbo coded FDPIM are simulated based on three kinds of analytical models under weak turbulence channel. The results show that packet error rate of uncoded FDPIM is inferior to that of uncoded DPIM. However, FDPIM is easy to be implemented and easy to be combined, with Turbo code for soft-decision because of its fixed length. Besides, the introduction of Turbo code in this modulation can decrease the average power about 10 dBm, which means that it can improve the error performance of the system effectively.     

A method for calculating the center position of collimated light

It is a key problem to accurately calculate beam spots＇ center of measuring the warp by using a collimated laser. A new method, named double geometrical center method （DGCM）, is put forward for the first time. In this method, a plane wave perpendicularly irradiates an aperture stop, and a charge couple device （CCD） is employed to receive the diffraction-beam spots, then the geometrical centers of the fast and the second diffraction-beam spots are calculated respectively, and their mean value is regarded as the center of datum beam. In face of such adverse instances as laser intension distributing defectively, part of the image being saturated, this method can still work well. What＇s more, this method can detect whether an unacceptable error exits in the courses of image receiving, processing and calculating. The experimental results indicate the precision of this method is high.     

Resolution Enhancement Technology:Integrated Process Solutions using OPC

DUV lithography, using the 248 nm wavelength, is a viable manufacturing option for devices with features at 130 nm and less. Given the low kl value of the lithography, integrated process development is a necessary method for achieving acceptable process latitude. The application of assist features for rule based OPC requires the simultaneous optimization of the mask, illumination optics and the resist.Described in this paper are the details involved in optimizing each of these aspects for line and space imaging.A reference pitch is first chosen to determine how the optics will be set. The ideal sigma setting is determined by a simple geometrically derived expression. The inner and outer machine settings are determined, in turn,with the simulation of a figure of merit. The maximum value of the response surface of this FOM occurs at the optimal sigma settings. Experimental confirmation of this is shown in the paper.Assist features are used to modify the aerial image of the more isolated images on the mask. The effect that the diffraction of the scattering bars (SBs) has on the image intensity distribution is explained. Rules for determining the size and placement of SBs are also given.Resist is optimized for use with off-axis illumination and assist features. A general explanation of the material' s effect is discussed along with the affect on the through-pitch bias. The paper culminates with the showing of the lithographic results from the fully optimized system.     

Cost-effective test solutions for smart card and other integrated flash applications

From its emergence in the late 1980s as a lower cost alternative to early EEPROM technologies, flash memory has evolved to higher densities and speedsand rapidly growing acceptance in mobile applications.In the process, flash memory devices have placed increased test requirements on manufacturers. Today, as flash device test grows in importance in China, manufacturers face growing pressure for reduced cost-oftest, increased throughput and greater return on investment for test equipment. At the same time, the move to integrated flash packages for contactless smart card applications adds a significant further challenge to manufacturers seeking rapid, low-cost test.