基于深度学习的文本自动摘要方案

针对自然语言处理（NLP）生成式自动摘要领域的语义理解不充分、摘要语句不通顺和摘要准确度不够高的问题，提出了一种新的生成式自动摘要解决方案，包括一种改进的词向量生成技术和一个生成式自动摘要模型。改进的词向量生成技术以Skip-Gram方法生成的词向量为基础，结合摘要的特点，引入词性、词频和逆文本频率三个词特征，有效地提高了词语的理解；而提出的Bi-MulRnn+生成式自动摘要模型以序列映射（seq2seq）与自编码器结构为基础，引入注意力机制、门控循环单元（GRU）结构、双向循环神经网络（BiRnn）、多层循环神经网络（MultiRnn）和集束搜索，提高了生成式摘要准确性与语句流畅度。基于大规模中文短文本摘要（LCSTS）数据集的实验结果表明，该方案能够有效地解决短文本生成式摘要问题，并在Rouge标准评价体系中表现良好，提高了摘要准确性与语句流畅度。     

Rotation,Translation and Scale Invariant Sign Word Recognition Using Deep Learning

Communication between people with disabilities and people who do not understand sign language is a growing social need and can be a tedious task. One of the main functions of sign language is to communicate with each other through hand gestures. Recognition of hand gestures has become an important challenge for the recognition of sign language. There are many existing models that can produce a good accuracy, but if the model test with rotated or translated images, they may face some difficulties to make good performance accuracy. To resolve these challenges of hand gesture recognition, we proposed a Rotation, Translation and Scale-invariant sign word recognition system using a convolutional neural network (CNN). We have followed three steps in our work: rotated, translated and scaled (RTS) version dataset generation, gesture segmentation, and sign word classification. Firstly, we have enlarged a benchmark dataset of 20 sign words by making different amounts of Rotation, Translation and Scale of the original images to create the RTS version dataset. Then we have applied the gesture segmentation technique. The segmentation consists of three levels, i) Otsu Thresholding with YCbCr, ii) Morphological analysis: dilation through opening morphology and iii) Watershed algorithm. Finally, our designed CNN model has been trained to classify the hand gesture as well as the sign word. Our model has been evaluated using the twenty sign word dataset, five sign word dataset and the RTS version of these datasets. We achieved 99.30% accuracy from the twenty sign word dataset evaluation, 99.10% accuracy from the RTS version of the twenty sign word evolution, 100% accuracy from the five sign word dataset evaluation, and 98.00% accuracy from the RTS version five sign word dataset evolution. Furthermore, the influence of our model exists in competitive results with state-of-the-art methods in sign word recognition.     

基于改进PVANet的实时小目标检测方法

现有目标检测算法主要以图像中的大目标作为研究对象,针对小目标的研究比较少且存在检测精确度低、无法满足实时性要求的问题,基于此,提出一种基于深度学习目标检测框架PVANet的实时小目标检测方法。首先,构建一个专用于小目标检测的基准数据集,它包含的目标在一幅图像中的占比非常小且存在截断、遮挡等干扰,可以更好地评估小目标检测方法的优劣;其次,结合区域建议网络（RPN）提出一种生成高质量小目标候选框的方法以提高算法的检测精确度和速度;选用step和inv两种新的学习率策略以改善模型性能,进一步提升检测精确度。在构建的小目标数据集上,相比原PVANet算法平均检测精确度提高了10.67%,速度提升了约30%。实验结果表明,该方法是一个有效的小目标检测算法,达到了实时检测的效果。     

分段卷积神经网络在文本情感分析中的应用

文本情感分析是当前网络舆情分析、产品评价、数据挖掘等领域的重要任务。由于当前网络数据的急剧增长,依靠人工设计特征或者传统的自然语言处理语法分析工具等进行分析,不但准确率不高而且费时费力。而传统的卷积神经网络模型均未考虑句子的结构信息,并且在训练时很容易发生过拟合。针对这两方面的不足,使用基于深度学习的卷积神经网络模型分析文本的情感倾向,采用分段池化的策略将句子结构考虑进来,分段提取句子不同结构的主要特征;并且引入Dropout算法以避免模型的过拟合和提升泛化能力。实验结果表明,分段池化策略和Dropout算法均有助于提升模型的性能,所提方法在中文酒店评价数据集上达到了91%的分类准确率,在斯坦福英文情感树库数据集五分类任务上达到了45.9%的准确率,较基线模型都有显著的提升。     

无池化层卷积神经网络的中文分词方法

在中文信息处理中,分词是一个十分常见且关键的任务。很多中文自然语言处理的任务都需要先进行分词,再根据分割后的单词完成后续任务。近来,越来越多的中文分词采用机器学习和深度学习方法。然而,大多数模型都不同程度的有模型过于复杂、过于依赖人工处理特征、对未登录词表现欠佳等缺陷。提出一种基于卷积神经网络(Convolutional Neural Networks,CNN)的中文分词模型——PCNN(Pure CNN)模型,该模型使用基于字向量上下文窗口的方式对字进行标签分类,具有结构简单、不依赖人工处理、稳定性好、准确率高等优点。考虑到分布式字向量本身的特性,在PCNN模型中不需要卷积的池化(Pooling)操作,卷积层提取的数据特征得到保留,模型训练速度获得较大提升。实验结果表明,在公开的数据集上,模型的准确率达到当前主流神经网络模型的表现水准,同时在对比实验中也验证了无池化层(Pooling Layer)的网络模型要优于有池化层的网络模型。     

融合语料库特征与图注意力网络的短文本分类方法

短文本分类是自然语言处理（NLP）中的重要研究问题,广泛应用于新闻分类、情感分析、评论分析等领域。针对短文本分类中存在的数据稀疏性问题,通过引入语料库的节点和边权值特征,基于图注意力网络（GAT）,提出了一个融合节点和边权值特征的图注意力网络NE-GAT。首先,针对每个语料库构建异构图,利用引力模型（GM）评估单词节点的重要性,并通过节点间的点互信息（PMI）获得边权重;其次,为每个句子构建文本级别图,并将节点重要性和边权重融入节点更新过程。实验结果表明,所提模型在测试集上的平均准确率达到了75.48%,优于用于文本分类的图卷积网络（Text-GCN）、TL-GNN、Text-ING等模型;相较原始GAT,所提模型的平均准确率提升了2.32个百分点,验证了其有效性。     

基于CNN-BLSTM-CRF模型的生物医学命名实体识别

命名实体识别是自然语言处理任务的重要步骤。近年来,不依赖人工特征的神经网络在新闻等通用领域命名实体识别方面表现出了很好的性能。然而在生物医学领域,许多实验表明基于领域知识的人工特征对于神经网络模型的结果影响很大。因此,如何在不依赖人工特征的情况下获得较好的生物医学命名实体识别性能是有待解决的问题。该文提出一种基于CNN-BLSTM-CRF的神经网络模型。首先利用卷积神经网络(CNN)训练出单词的具有形态特征的字符级向量,并从大规模背景语料训练中得到具有语义特征信息的词向量,然后将二者进行组合作为输入,再构建适合生物医学命名实体识别的BLSTM-CRF深层神经网络模型。实验结果表明,不依赖任何人工特征,该文方法在Biocreative Ⅱ GM和JNLPBA2004生物医学语料上都达到了目前最好的结果,F-值分别为89.09%和74.40%。     

基于细粒度词表示的命名实体识别研究

命名实体识别(NER)是自然语言处理中的一项基础任务,其性能的优劣极大地影响着关系抽取、语义角色标注等后续任务。传统的统计模型特征设计难度大、领域适应性差,一些神经网络模型则忽略了词本身所具有的形态学信息。针对上述问题,该文构建了一种基于细粒度词表示的端到端模型(Finger-BiLSTM-CRF)来进行命名实体识别任务。该文首先提出一种基于注意力机制的字符级词表示模型Finger来融合形态学信息和单词的字符信息,然后将Finger与BiLSTM-CRF模型联合进行实体识别,最终该方法以端到端、无任何特征工程的方式在CoNLL 2003 数据集上取得了F₁为91.09%的结果。实验表明,该文设计的Finger模型显著提升NER系统的召回率,从而使得模型的识别能力显著提升。     

基于word embedding和CNN的情感分类模型

尝试将word embedding和卷积神经网络（CNN）相结合来解决情感分类问题。首先,利用Skip-Gram模型训练出数据集中每个词的word embedding,然后将每条样本中出现的word embedding组合为二维特征矩阵作为卷积神经网络的输入;此外,每次迭代训练过程中,输入特征也作为参数进行更新。其次,设计了一种具有3种不同大小卷积核的神经网络结构,从而完成多种局部抽象特征的自动提取过程。与传统机器学习方法相比,所提出的基于word embedding和CNN的情感分类模型成功将分类正确率提升了5.04%。     

Deep Learning Enabled Social Media Recommendation Based on User Comments

Nowadays, review systems have been developed with social media Recommendation systems (RS). Although research on RS social media is increasing year by year, the comprehensive literature review and classification of this RS research is limited and needs to be improved. The previous method did not find any user reviews within a time, so it gets poor accuracy and doesn’t filter the irrelevant comments efficiently. The Recursive Neural Network-based Trust Recommender System (RNN-TRS) is proposed to overcome this method’s problem. So it is efficient to analyse the trust comment and remove the irrelevant sentence appropriately. The first step is to collect the data based on the transactional reviews of social media. The second step is pre-processing using Imbalanced Collaborative Filtering (ICF) to remove the null values from the dataset. Extract the features from the pre-processing step using the Maximum Support Grade Scale (MSGS) to extract the maximum number of scaling features in the dataset and grade the weights (length, count, etc.). In the Extracting features for Training and testing method before that in the feature weights evaluating the softmax activation function for calculating the average weights of the features. Finally, In the classification method, the Recursive Neural Network-based Trust Recommender System (RNN-TRS) for User reviews based on the Positive and negative scores is analysed by the system. The simulation results improve the predicting accuracy and reduce time complexity better than previous methods.     

基于CNN‐Head Transformer编码器的中文命名实体识别

基于多头自注意力机制的Transformer作为主流特征提取器在多种自然语言处理任务中取得了优异表现,但应用于命名实体识别任务时存在一字多词、增加额外存储与词典匹配时间等问题。提出一种CNN-Head Transformer编码器（CHTE）模型,在未使用外部词典和分词工具的基础上,通过自注意力机制捕获全局语义信息,利用不同窗口大小的CNN获取Transformer中6个注意力头的Value向量,使CHTE模型在保留全局语义信息的同时增强局部特征和潜在词信息表示,并且应用自适应的门控残差连接融合当前层和子层特征,提升了Transformer在命名实体识别领域的性能表现。在Weibo和Resume数据集上的实验结果表明,CHTE模型的F1值相比于融合词典信息的Lattice LSTM和FLAT模型分别提升了3.77、2.24和1.30、0.31个百分点,具有更高的中文命名实体识别准确性。     

基于词向量和卷积神经网络的垃圾短信识别方法

对垃圾短信进行过滤识别研究具有重要的社会价值和时代背景意义。针对传统的人工设计短信特征选择方法中存在数据稀疏、特征信息共现不足和特征提取困难的问题,提出一种基于词向量和卷积神经网络（CNN）的垃圾短信识别方法。首先,使用word2vec的skip-gram模型根据维基中文语料库训练出短信数据集中每个词的词向量,并将每条短信中各个词组所对应的词向量组成表示短信的二维特征矩阵;然后,把特征矩阵作为卷积神经网络的输入,通过卷积层的不同尺度卷积核提取多尺度短信特征,以及利用1-max pooling池化策略得到局部最优特征;最后,将局部最优特征组成融合特征向量放入softmax分类器中得出分类结果。在10万条短信数据上进行的实验结果表明,在特征提取方式相同的情况下,基于卷积神经网络模型的识别准确率能够达到99.5%,比传统的机器学习模型提高了2.4%~5.1%,且各模型的识别准确率均保持在94%以上。     

通用深度学习语言模型的隐私风险评估

近年来,自然语言处理领域涌现出多种基于Transformer网络结构的通用深度学习语言模型,简称“通用语言模型(general-purpose language models, GPLMs)”,包括Google提出的BERT(bidirectional encoder representation from transformers)模型等,已在多个标准数据集和多项重要自然语言处理任务上刷新了最优基线指标,并已逐渐在商业场景中得到应用.尽管其具有很好的泛用性和性能表现,在实际部署场景中,通用语言模型的安全性却鲜为研究者所重视.近年有研究工作指出,如果攻击者利用中间人攻击或作为半诚实(honest-but-curious)服务提供方截获用户输入文本经由通用语言模型计算产生的文本特征,它将以较高的准确度推测原始文本中是否包含特定敏感词.然而,该工作仅采用了特定敏感词存在与否这一单一敏感信息窃取任务,依赖一些较为严格的攻击假设,且未涉及除英语外其他语种的使用场景.为解决上述问题,提出1条针对通用文本特征的隐私窃取链,从更多维度评估通用语言模型使用中潜在的隐私风险.实验结果表明：仅根据通用语言模型提取出的文本表征,攻击者能以近100％的准确度推断其模型来源,以超70％的准确度推断其原始文本长度,最终推断出最有可能出现的敏感词列表,以重建原始文本的敏感语义.此外,额外针对3种典型的中文预训练通用语言模型开展了相应的隐私窃取风险评估,评估结果表明中文通用语言模型同样存在着不可忽视的隐私风险.     

A computational method for solving two‐dimensional nonlinear variable‐order fractional optimal control problems

In this paper, a new class of two‐dimensional nonlinear variable‐order fractional optimal control problems (V‐OFOCPs) is introduced where the variable‐order fractional derivative is defined in the Caputo type. The general procedure for solving theses systems is expanding the state variable and the control variable based on the Legendre cardinal functions in the matrix form. Hence, we derive their operational matrix of derivative (OMD) and operational matrix of variable‐order fractional derivative (OMV‐OFD). More significantly, some properties of these basis functions are proved to be exploited in our approach. Using these achieved results, we simply expand the matrix form of the nonlinear performance index in terms of the Legendre cardinal functions and subsequently convert it to an algebraic equation. We emphasize that it is a valuable advantage of applying cardinal functions in approximation theory. Then, we implement the OMD and the OMV‐OFD of the Legendre cardinal functions to transform the variable‐order fractional dynamical system to a system of algebraic equations. Next, the method of constrained extremum is applied to adjoin the constraint equations including the given dynamical system and the initial‐boundary conditions to the performance index by a set of undetermined Lagrange multipliers. Finally, the necessary conditions of the optimality are derived as a system of nonlinear algebraic equations including the unknown coefficients of the state variable, the control variable and the Lagrange multipliers. The applicability and efficiency of the proposed approach are investigated through the various types of test problems.     

A new Wavelet Method for Variable‐Order Fractional Optimal Control Problems

In this paper, a new computational method based on the Legendre wavelets (LWs) is proposed for solving a class of variable‐order fractional optimal control problems (V‐FOCPs). To do this, a new operational matrix of variable‐order fractional integration (OMV‐FI) in the Riemann‐Liouville sense for the LWs is derived and used to obtain an approximate solution for the problem under study. Along the way the hat functions (HFs) are introduced and employed to derive a general procedure to compute this matrix. In the proposed method, the variable‐order fractional dynamical system is transformed to an equivalent variable‐order fractional integro‐differential dynamical system, at first. Then, the highest integer order of the derivative of the state variable and the control variable are expanded by the LWs with unknown coefficients. Next, the OMV‐FI in the the Riemann‐Liouville sense together with some properties of the LWs are employed to achieve a nonlinear algebraic equation in place of the performance index and a nonlinear system of algebraic equations in place of the dynamical system in terms of the unknown coefficients. Finally, the method of constrained extremum is applied which consists of adjoining the constraint equations derived from the given dynamical system to the performance index by a set of undetermined Lagrange multipliers. As a result, the necessary conditions of optimality are derived as a system of algebraic equations in the unknown coefficients of the state variable, control variable and Lagrange multipliers. Furthermore, the efficiency and accuracy of the proposed method are demonstrated for some concrete examples. The obtained results show that the proposed method is very efficient and accurate.     

Learning-based word spotting system for Arabic handwritten documents

The retrieval of information from scanned handwritten documents is becoming vital with the rapid increase of digitized documents, and word spotting systems have been developed to search for words within documents. These systems can be either template matching algorithms or learning based. This paper presents a coherent learning based Arabic handwritten word spotting system which can adapt to the nature of Arabic handwriting, which can have no clear boundaries between words. Consequently, the system recognizes Pieces of Arabic Words (PAWs), then re-constructs and spots words using language models. The proposed system produced promising result for Arabic handwritten word spotting when tested on the CENPARMI Arabic documents database.     

融合基于语言模型的词嵌入和多尺度卷积神经网络的情感分析

针对Word2Vec、GloVe等词嵌入技术对多义词只能产生单一语义向量的问题,提出一种融合基于语言模型的词嵌入（ELMo）和多尺度卷积神经网络（MSCNN）的情感分析模型。首先,该模型利用ELMo学习预训练语料,生成上下文相关的词向量;相较于传统词嵌入技术,ELMo利用双向长短程记忆（LSTM）网络融合词语本身特征和词语上下文特征,能够精确表示多义词的多个不同语义;此外,该模型使用预训练的中文字符向量初始化ELMo的嵌入层,相对于随机初始化,该方法可加快模型的训练速度,提高训练精度;然后,该模型利用多尺度卷积神经网络,对词向量的特征进行二次抽取,并进行特征融合,生成句子的整体语义表示;最后,经过softmax激励函数实现文本情感倾向的分类。实验在公开的酒店评论和NLPCC2014 task2两个数据集上进行,实验结果表明,在酒店评论数据集上与基于注意力的双向LSTM模型相比,该模型正确率提升了1.08个百分点,在NLPCC2014 task2数据集上与LSTM和卷积神经网络（CNN）的混合模型相比,该模型正确率提升了2.16个百分点,证明了所提方法的有效性。     

基于BIG-WFCHI的微博信息关键特征选择方法

特征选择是用机器学习方法提高转发预测精度和效率的关键步骤,其前提是特征提取.目前,特征选择中常用的方法有信息增益(Information Gain,IG)、互信息和卡方检验(CHI-square test,CHI)等,传统特征选择方法中出现低频词引起的信息增益和卡方检验的负相关、干扰计算等问题,导致分类准确率不高.本文...     

基于注意力和字嵌入的中文医疗问答匹配方法

针对当前的分词工具在中文医疗领域无法有效切分出所有医学术语,且特征工程需消耗大量人力成本的问题,提出了一种基于注意力机制和字嵌入的多尺度卷积神经网络建模方法。该方法使用字嵌入结合多尺度卷积神经网络用以提取问题句子和答案句子不同尺度的上下文信息,并引入注意力机制来强调问题和答案句子之间的相互影响,该方法能有效学习问题句子和正确答案句子之间的语义关系。由于中文医疗领域问答匹配任务没有标准的评测数据集,因此使用公开可用的中文医疗问答数据集（cMedQA）进行评测,实验结果表明该方法优于词匹配、字匹配和双向长短时记忆神经网络（BiLSTM）建模方法,并且Top-1准确率为65.43%。     

基于迁移孪生非负矩阵分解的静脉识别算法

针对某一静脉图像数据集下获得的识别算法对于其他数据集缺少普适性的问题,提出了一种具有迁移性的孪生非负矩阵分解(NMF)模型.首先,通过采用两个结构相同且参数共享的NMF模型实现了对源数据集中带有相同标签静脉图像的有监督学习;然后,通过使用最大均值差异约束降低了不同数据集之间静脉特征的差异性,即将源数据集中的知识迁移至目...