基于转移神经网络的中文AMR解析

抽象语义表示(abstract meaning representation, AMR)是一种领域无关的句子语义表示方法,它将一个句子的语义抽象为一个单根有向无环图,AMR解析旨在将句子解析为对应的AMR图。目前,中文AMR研究仍然处于起步阶段。该文结合中文AMR特性,采用基于转移神经网络的方法对中文AMR解析问题展开了试验性研究。首先,实现了一个基于转移解码方法的增量式中文AMR解析神经网络基线系统;然后,通过引入依存路径语义关系表示学习和上下文相关词语语义表示学习,丰富了特征的表示;最后,模型中应用序列化标注的模型实现AMR概念识别,优化了AMR概念识别效果。实验结果表明,该模型在中文AMR解析任务中达到了0.61的Smatch F1值,明显优于基线系统。     

融入句法感知表示进行句法增强的语义解析

在自然语言处理（NLP）中,句法信息是完整句子中词汇与词汇之间的句法结构关系或者依存关系,是一种重要且有效的参考信息。语义解析任务是将自然语言语句直接转化成语义完整的、计算机可执行的语言。在以往的语义解析研究中,少有采用输入源的句法信息来提高端到端语义解析效率的工作。为了进一步提高端到端语义解析模型的准确率和效率,提出一种利用输入端句法依存关系信息来提高模型效率的语义解析方法。该方法的基本思路是先对一个端到端的依存关系解析器进行预训练;然后将该解析器的中间表示作为句法感知表示,与原有的字词嵌入表示拼接到一起以产生新的输入嵌入表示,并将得到的输入嵌入表示用于端到端语义解析模型;最后采用转导融合学习方式进行模型融合。实验对比了所提模型和基准模型Transformer以及过去十年的相关工作。实验结果表明,在ATIS、GEO、JOBS数据集上,融入依存句法信息感知表示以及转导融合学习的语义解析模型分别实现了89.1%、90.7%、91.4%的最佳准确率,全面超过了Transformer,验证了引入句法依存关系信息的有效性。     

一种基于词对齐的中文深层语义解析模型

语义解析是指将自然语言句子转化成便于机器理解和推理的意义形式。近年来英文语义解析的研究取得了很大进展。然而,中文语义解析的相关工作则相对较少。中文和英文之间存在一定的差异,适用于英文的语义解析方法不一定适合中文。因此,针对中文的语言特点,提出一种基于词对齐的中文语义解析方法,将中文句子转化成其相应的意义表示看作是一个机器翻译的过程。首先将英文语义解析方法中常用的训练数据集GEOQUERY转化成中文数据集,数据集中每条训练数据包括一个中文句子及其正确的意义表示。然后利用词对齐模型来获取由中文自然语言字符串及其相应的意义表示所组成的双语词典。最后通过学习一个概率估计模型来确定最终的语义解析模型。实验结果表明,WACSP有较高的精确度和覆盖率。     

基于最大熵模型的汉语问句语义组块分析

问句分析是问答系统的关键,为降低问句完整语法分析的复杂度,该文应用浅层句法分析理论,采用问句语义组块方式来分析问句。以“知网”知识库为基础,提取和定义了表达汉语问句的6种语义块,定义了语义组块最大熵模型的特征表示,通过最大熵原理实现了语义组块特征抽取和特征选取学习算法,并以模型为基础实现了真实问句的语义块的标注,从而为在语义层面上理解汉语问句奠定了基础。实验结果说明最大熵模型应用于汉语问句语义组块分析具有较好的效果。     

半监督学习和主动学习相结合的浅层语义分析

语义分析是基于内容的文本挖掘领域的重要技术和研究难点。有监督机器学习方法受限于标注语料的规模,在小规模标注样本中难以获取较高性能。本文面向浅层语义分析任务,采用一种新颖的半监督学习方法——直推式支持向量机,并结合其训练特点提出了基于主动学习的样本优化策略。实验表明,本文提出的浅层语义分析方法通过整合主动学习与半监督学习,在小规模标注样本环境中取得了良好的学习效果。     

基于补全信息的篇章级神经机器翻译

对于句子级别的神经机器翻译,由于不考虑句子所处的上下文信息,往往存在句子语义表示不完整的问题。该文通过依存句法分析,对篇章中的每句话提取有效信息,再将提取出的信息,补全到源端句子中,使得句子的语义表示更加完整。该文在汉语-英语语言对上进行了实验,并针对篇章语料稀少的问题,提出了在大规模句子级别的平行语料上的训练方法。相比于基准系统,该文提出的方法获得了1.47个BLEU值的提高。实验表明,基于补全信息的篇章级神经机器翻译,可以有效地解决句子级别神经机器翻译语义表示不完整的问题。     

REVERSING A LEXICALLY BASED PARSER FOR GENERATION

It is practically advantageous and theoretically meaningful to base both a natural language generator and a parser on the same linguistic data. Neutrality of the representation can be advocated for all unification-based grammars. Yet the basic process in generation can also have several points in common with parsing. Specifically, not only is a unification-based approach proposed that produces from the initial semantic representation the sentence together with a functional structure but also the process is based on an adaptation of the concept of chart, which we call bi-chart (bidimensional chart). The chart is cast over the semantic representation and not over the string but has the same qualities as the classic one. The other relevant point here is that most of the data used in the process are distributed throughout the lexicon, and this results in a very flexible system.     

藏文句子语义块识别方法

语义理解是自然语言理解的一项关键任务,传统上采用以语法为中心的词法和句法分析等技术来解析句义。该文提出了一种以语义块分析藏文句义的新方法,其中藏文语义块识别通过采用Bi-LSTM和ID-CNN 两种神经网络构架对该任务进行建模和对比分析。经实验,上述的两种模型在测试数据集上取得了良好的性能表现, F₁值平均分别为89%和92%。这种语义块分析和识别技术能够较好地替代词义消歧和语义角色标注等工作。     

基于密集连接卷积神经网络的远程监督关系抽取

密集连接卷积神经网络(DenseNet)是一种新型深度卷积神经网络架构,通过建立不同层间的连接关系,来确保网络层与层间最大程度的信息传输。在文本远程监督关系抽取任务中,针对现有神经网络方法使用浅层网络提取特征的局限,设计了一种基于密集连接方式的深度卷积神经网络模型。该模型采用五层卷积神经网络构成的密集连接模块和最大池化层作为句子编码器,通过合并不同层次的词法、句法和语义特征,来帮助网络学习特征,从而获取输入语句更丰富的语义信息,同时减轻深度神经网络的梯度消失现象,使得网络对自然语言的表征能力更强。模型在NYT-Freebase数据集上的平均准确率达到了82.5%,PR曲线面积达到了0.43。实验结果表明,该模型能够有效利用特征,并提高远程监督关系抽取的准确率。     

融合源端句法和语义角色信息的AMR解析

序列到序列(seq2seq)的框架可以应用到抽象语义表示(AMR)解析任务中,把AMR解析当作一个从源端句子到目标端AMR图的翻译任务。然而,以前的工作通常把源端句子表示为一个单词序列,忽略了句子内部潜藏的句法和语义角色信息。基于seq2seq框架,该文提出了一个直接而有效的融合句法和语义角色信息的AMR解析方法。实验结果表明,该文的方法在AMR英文标准数据集上取得了6.7%的显著提升。最后,该文从多个角度深入分析了源端的句法和语义角色信息是如何对AMR解析提供帮助的。分析表明,词性信息和subword技术对AMR解析性能提升的贡献最大,上层句法和语义角色信息次之。     

Abstract meaning representation for legal documents: an empirical research on a human-annotated dataset

Natural language processing techniques contribute more and more in analyzing legal documents recently, which supports the implementation of laws and rules using computers. Previous approaches in representing a legal sentence often based on logical patterns that illustrate the relations between concepts in the sentence, often consist of multiple words. Those representations cause the lack of semantic information at the word level. In our work, we aim to tackle such shortcomings by representing legal texts in the form of abstract meaning representation (AMR), a graph-based semantic representation that gains lots of polarity in NLP community recently. We present our study in AMR Parsing (producing AMR from natural language) and AMR-to-text Generation (producing natural language from AMR) specifically for legal domain. We also introduce JCivilCode, a human-annotated legal AMR dataset which was created and verified by a group of linguistic and legal experts. We conduct an empirical evaluation of various approaches in parsing and generating AMR on our own dataset and show the current challenges. Based on our observation, we propose our domain adaptation method applying in the training phase and decoding phase of a neural AMR-to-text generation model. Our method improves the quality of text generated from AMR graph compared to the baseline model. (This work is extended from our two previous papers: “An Empirical Evaluation of AMR Parsing for Legal Documents”, published in the Twelfth International Workshop on Juris-informatics (JURISIN) 2018; and “Legal Text Generation from Abstract Meaning Representation”, published in the 32nd International Conference on Legal Knowledge and Information Systems (JURIX) 2019.).

     

基于朴素贝叶斯与潜在狄利克雷分布相结合的情感分析

针对情感分析需要大量人工标注语料的难点,提出了一种面向无指导情感分析的层次性生成模型。该模型将朴素贝叶斯(NB)模型和潜在狄利克雷分布(LDA)相结合,仅仅需要合适的情感词典,不需要篇章级别和句子级别的标注信息即可同时对网络评论的篇章级别和句子级别的情感倾向进行分析。该模型假设每个句子而不是每个单词拥有一个潜在的情感变量;然后,该情感变量再以朴素贝叶斯的方式生成一系列独立的特征。在该模型中,朴素贝叶斯假设的引入使得该模型可以结合自然语言处理(NLP)相关的技术,例如依存分析、句法分析等,用以提高无指导情感分析的性能。在两个情感语料数据集上的实验结果显示,该模型能够自动推导出篇章级别和句子级别的情感极性,该模型的正确率显著优于其他无指导的方法,甚至接近部分半指导或有指导的研究方法。     

Objectness Region Enhancement Networks for Scene Parsing

Semantic segmentation has recently witnessed rapid progress, but existing methods only focus on identifying objects or instances. In this work, we aim to address the task of semantic understanding of scenes with deep learning. Different from many existing methods, our method focuses on putting forward some techniques to improve the existing algorithms, rather than to propose a whole new framework. Objectness enhancement is the first effective technique. It exploits the detection module to produce object region proposals with category probability, and these regions are used to weight the parsing feature map directly. “Extra background” category, as a specific category, is often attached to the category space for improving parsing result in semantic and instance segmentation tasks. In scene parsing tasks, extra background category is still beneficial to improve the model in training. However, some pixels may be assigned into this nonexistent category in inference. Black-hole filling technique is proposed to avoid the incorrect classification. For verifying these two techniques, we integrate them into a parsing framework for generating parsing result. We call this unified framework as Objectness Enhancement Network (OENet). Compared with previous work, our proposed OENet system effectively improves the performance over the original model on SceneParse150 scene parsing dataset, reaching 38.4 mIoU (mean intersectionover-union) and 77.9% accuracy in the validation set without assembling multiple models. Its effectiveness is also verified on the Cityscapes dataset.     

Knowledge-guided unsupervised rhetorical parsing for text summarization

Automatic text summarization (ATS) has recently achieved impressive performance thanks to recent advances in deep learning and the availability of large-scale corpora. However, there is still no guarantee that the generated summaries are grammatical, concise, and convey all salient information as the original documents have. To make the summarization results more faithful, this paper presents an unsupervised approach that combines rhetorical structure theory, deep neural model, and domain knowledge concern for ATS. This architecture mainly contains three components: domain knowledge base construction based on representation learning, the attentional encoder–decoder model for rhetorical parsing, and subroutine-based model for text summarization. Domain knowledge can be effectively used for unsupervised rhetorical parsing thus rhetorical structure trees for each document can be derived. In the unsupervised rhetorical parsing module, the idea of translation was adopted to alleviate the problem of data scarcity. The subroutine-based summarization model purely depends on the derived rhetorical structure trees and can generate content-balanced results. To evaluate the summary results without golden standard, we proposed an unsupervised evaluation metric, whose hyper-parameters were tuned by supervised learning. Experimental results show that, on a large-scale Chinese dataset, our proposed approach can obtain comparable performances compared with existing methods.     

基于语义解析的中文GIS自然语言接口实现研究

该文对基于语义解析的中文地理信息系统(GIS)自然语言接口实现技术与方法进行了探索性的研究。首先,我们针对一个具体GIS应用领域设计和开发了一种函数式的形式化意义表示语言GISQL和一个中文语义解析标注语料库;然后,我们通过引入混合树作为隐变量用于构造输入句子与输出表示结构之间的对应关系,提出了一种基于含隐变量的感知器模型的语义解析算法。在开发的中文语义解析标注语料库上的实验结果显示,该文提出的语义解析算法的F1值达到了90.67％,明显优于baseline系统。更重要的是,该文的研究证明了基于语义解析方法实现中文GIS的自然语言接口是一种有效可行的途径。     

结合依存句法分析与交互注意力机制的隐式方面提取

隐式方面提取对于提升细粒度情感分析的准确性具有重要意义,然而现有隐式方面提取技术在处理大规模数据时泛化能力不强。为此,提出结合依存句法分析与交互注意力机制的隐式方面提取模型。首先利用预训练语言模型BERT生成文本的初始表征,然后传递给依存句法引导的自注意力层再次处理,再将两次处理的结果经交互注意力机制进一步提取特征,最终用分类器判断句子所属的隐式方面类别。与基线BERT及其他深度神经网络模型对比,所提模型在增强的SemEval隐式方面数据集上取得了更高的F₁与AUC值,证明了模型的有效性。     

基于交互注意力图卷积网络的方面情感分类

基于方面的情感分类任务旨在识别句子中给定方面词的情感倾向性。以往的方法大多基于长短时记忆网络和注意力机制,这种做法在很大程度上仅依赖于建模句子中的方面词与其上下文的语义相关性,但忽略了句中的语法信息。针对这种缺陷,提出了一种交互注意力的图卷积网络,同时建模了句中单词的语义相关性和语法相关性。首先使用双向长短时记忆网络来学习句子的词序关系,捕捉句中上下文的语义信息;其次引入位置信息后,通过图卷积网络来学习句中的语法信息;然后通过一种掩码机制提取方面词;最后使用交互注意力机制,交互计算特定方面的上下文表示,并将其作为最后的分类特征。通过这种优势互补的设计,该模型可以很好地获得聚合了目标方面信息的上下文表示,并有助于情感分类。实验结果表明,该模型在多个数据集上都获得了优秀的效果。与未考虑语法信息的Bi-IAN模型相比,该模型在所有数据集上的结果均优于Bi-IAN模型,尤其在餐厅领域的REST14,REST15和REST16数据集上,该模型的F1值较Bi-IAN模型分别提高了4.17%,7.98%和8.03%;与同样考虑了语义信息和语法信息的ASGCN模型相比,该模型的F1值在除了LAP14数据集外的其他数据集上均优于ASGCN模型,尤其在餐厅领域的REST14,REST15和REST16数据集上,该模型的F1值较ASGCN模型分别提高了2.05%,1.66%和2.77%。     

RECOGNITION CAN BE HARDER THAN PARSING

The work presented here attempts to bring out some fundamental concepts that underlie some known parsing algorithms, usually called chart or dynamic programming parsers, in the hope of guiding the design of similar algorithms for other formalisms that could be considered for describing the "surface" syntax of languages. The key idea is that chart parsing is essentially equivalent to a simple construction of the intersection of the language (represented by its grammar) with a regular set containing only the input sentence to be parsed (represented by a finite state machine). The resulting grammar for that intersection is precisely what is usually called a shared forest: it represents all parses of a syntactically ambiguous sentence. Since most techniques for processing ill-formed input can be modeled by considering a nonsingleton regular set of input sentences, we can expect to generalize these ill-formed input processing techniques to all parsers describable with our approach.     

基于句子跨度的哈萨克语句法分析研究

由于目前哈萨克语句法分析准确率较低并缺乏基于神经网络的哈萨克语句法分析的相关研究,针对哈萨克语短语结构的句法分析,使用基于移进—归约的方法,采用在栈中存储句子跨度而不是部分树结构,从而在进行句法树解析时不需要对句法树进行二叉化。该研究在句子特征提取时使用双向LSTM对句子跨度特征进行提取,得到句子跨度在整个句子上下文中信息,再使用多层感知机对句法分析模型进行训练,最后在解码时使用动态规划选取最优句法分析结果;最终使得哈萨克语短语句法分析准确率达到了76.92%。研究成果对哈萨克语句法分析准确率有了进一步的提高,并为后续的哈萨克语机器翻译及语义分析奠定良好的基础。     

基于语义理解和机器学习的混合的中文文本情感分类算法框架

快速、准确和全面地从大量互联网文本信息中定位情感倾向是当前大数据技术领域面临的一大挑战.文本情感分类方法大致分为基于语义理解和基于有监督的机器学习两类.语义理解处理情感分类的优势在于其对不同领域的文本都可以进行情感分类,但容易受到中文存在的不同句式及搭配的影响,分类精度不高.有监督的机器学习虽然能够达到比较高的情感分类精度,但在一个领域方面得到较高分类能力的分类器不适应新领域的情感分类.在使用信息增益对高维文本做特征降维的基础上,将优化的语义理解和机器学习相结合,设计了一种新的混合语义理解的机器学习中文情感分类算法框架.基于该框架的多组对比实验验证了文本信息在不同领域中高且稳定的分类精度.