【作者】 陈斌；

【导师】 周勇；

【作者基本信息】 中国矿业大学 ， 计算机应用技术， 2019， 硕士

【摘要】 近年来,随着互联网的爆炸式发展,海量的数据通过文本数字化的形式呈现出来。面对信息爆炸的时代,如何从无结构的文本数据中提取出人们关注度高的热点信息成为自然语言处理当前的主要研究方向。信息抽取技术被作为这一问题的解决方案受到研究人员的广泛关注。其中,事件抽取是信息抽取技术的重要环节,也是信息抽取领域中最具挑战性的任务之一。事件抽取的目的是从无结构的文本数据中抽取出描述事件的关键元素,并将事件结构化的展示出来。当前事件抽取技术主要是基于语料的有监督学习,对高质量的语料标注依赖性较强。事件抽取从抽取的过程上可以划分为事件触发词抽取和事件要素抽取两个阶段。事件抽取技术的研究对知识图谱、舆情分析,自动文摘和机器翻译等自然语言处理任务具有重要的指导意义。传统事件抽取方法大多存在关键特征提取不足,容易忽略上下文语境信息等问题。为了解决上述问题,本文在长短期记忆网络的基础上,引入卷积操作和注意力机制,进行了一系列事件抽取相关的研究。本文的主要研究内容包括:1、本文提出了基于卷积长短期记忆网络模型的事件触发词抽取方法。在事件触发词的特征选择上,采用词向量和位置向量作为事件句的文本向量化表达。为了解决传统事件抽取方法中一词多义现象造成的事件分类错误问题,利用长短期记忆网络提取句子级别特征,卷积操作提取词级别特征,而后将两种特征结合起来进入输出层进行事件触发词抽取结果的预测。最后,除了标准的事件触发词抽取结果外,本文还在模糊触发词抽取任务上验证了模型的有效性。2、本文以注意力机制为核心,在长短期记忆网络模型的基础上,搭建了基于Encoder-Decoder框架的事件要素抽取模型。在特征选取上,除了词向量与位置特征外还选取了事件触发词类型特征作为文本向量化的表示。在事件要素抽取模型的构建上,注意力层能够有效表现出事件触发词和其它事件要素对候选事件要素的影响,解决了当一个事件句中包含多个事件时,事件要素角色容易识别混乱的问题,从而提高了事件要素的抽取效率。3、本文将事件抽取模型应用在新闻舆情领域,实现了新闻舆情事件抽取原型系统。新闻舆情事件抽取原型系统在具体的事件触发词和事件要素两个阶段中,分别采用基于卷积长短期记忆网络模型和基于注意力机制的长短期记忆网络模型,以流水线的方式完成事件抽取的全部过程。为了解决系统在实际应用过程中训练语料规模不足、无法适应新领域的问题,本文设计了增量学习框架,增强了系统的健壮性和普遍适用性。在事件具体展示的实现上,本文除了实现事件的结构化展示,还从时间和空间的不同维度上设计了事件时间轴和事件地图功能,使新闻舆情事件的记录有序化、系统化和完整化。更多还原

【Abstract】 In recent years,with the rapid development of the Internet,a large amount of information has been presented to people in the form of electronic text.Facing the age of information explosion,how to extract interesting information from unstructured text has become one of the research hotspots,and information extraction technology came into being at the historic moment.Event extraction is an important part of information extraction technology and one of the most challenging tasks in the field of information extraction.The purpose of event extraction is to extract key elements describing the event from unstructured text data and show them structurally.The current event extraction technology is mainly based on the supervised method,which is highly dependent on high-quality corpus annotation.Event extraction can be divided into two stages: event trigger extraction and event argument extraction.The research of event extraction technology has important guiding significance for natural language processing tasks such as knowledge mapping,public opinion analysis,automatic summarization and machine translation.Most of the traditional event extraction methods have insufficient key feature extraction,and it is easy to ignore the context information.In order to solve the above problems,based on the long short-term memory network,this paper introduces the convolution operation and attention mechanism,and conducts a series of events related research.The main research contents of this paper include:1、This paper proposes an event trigger extraction method based on convolutional long short-term memory network.The word embedding and position feature are selected as the representation of text vectorization.Word embedding adopts Skip-gram model,and the position feature adopts a discretized 5-dimensional vector.To solve the problem of event classification error caused by polysemy in the traditional event extraction method,the long short-term memory neural network is used to extract the sentence-level features,the convolution operation is used to extract the lexical-level features,and the two features are combined into the output layer to predict the results of event trigger extraction.In the model training stage,the cross entropy cost function is used as the loss function,and the gradient descent method uses the Adam optimization algorithm.Except the standard event trigger extraction results,this paper also verifies the validity of the model on the fuzzy event trigger extraction.2、Based on the long short-term memory network,this paper introduces the attention mechanism and builds the event argument extraction model based on the Encoder-Decoder framework.In the feature selection,in addition to the word embedding and the position feature,the event trigger word type feature is selected as the representation of the text vectorization.In the construction of the event argument extraction model,the attention layer can effectively show the impact of event trigger words and other event arguments on the candidate event arguments.This solves the problem that when an event sentence contains multiple events,the event argument role is easy to identify confusion,which improves the efficiency of extracting event arguments.3、This paper applies the event extraction model to the field of news public opinion,and realizes the prototype news public opinion event extraction system.The prototype news public opinion event extraction system uses the Conv-BiLSTM model and the AB-BiLSTM model in the two stages of specific event trigger and event argument extraction,and completes the event extraction in a pipelined approach.In order to solve the problem that the training corpus is insufficient in the actual application process and cannot adapt to the new field,this paper designs an incremental learning framework,which enhances the robustness and universal applicability of the system.In the implementation of the event specific display,this paper also designed the event timeline function and event map function from different dimensions of time and space,so that the prototype news public opinion event extraction system are ordered,systematic and complete.更多还原