节点文献
面向中文文本的事件抽取方法研究
【作者】 张伟；
【导师】 王永利；
【作者基本信息】 南京理工大学 ， 智能计算与系统， 2021， 硕士
【摘要】 随着计算机技术和互联网规模的飞速发展,如何从日益增长的海量网络信息中抽取有用的信息,并通过结构化文本的形式呈现显得尤为重要。信息抽取技术作为这一问题的解决方案而备受关注。其中,事件抽取是信息抽取领域的一个重要研究方向,也是信息抽取领域中最具挑战性任务之一。事件抽取就是从由自然语言描述的非结构化文本中抽取用户关心和感兴趣的事件,然后用结构化的数据形式进行保存和展示,方便用户快速获取事件的核心信息,掌握事件对应的进展情况。事件抽取的任务从其处理流程中可以划分为两个子任务,分别是对事件触发词的抽取和事件元素的抽取。传统的事件抽取方法中存在易忽略上下文语境信息、文本关键特征提取不充分等问题。为了解决上述问题,本文使用深度学习的方法对中文本文的事件抽取展开了研究,具体的研究工作内容如下:1.针对中文事件类型的识别与分类问题,提出了一种基于循环神经网络的中文事件触发词检测方法。本文利用BERT模型对中文语料库进行预训练得到BERT预训练词向量,利用LTP工具对中文语句进行分析处理,获取词汇的词性向量、语义依存向量和句法分析向量,将这四种文本向量进行拼接作为Bi LSTM网络层的输入。经过Bi LSTM网络层计算得到文本的信息特征表示,经过CRF序列标注层处理后,最后得到事件触发词的检测结果。2.针对中文事件元素的识别与分类问题,提出了一种结合事件触发词信息与注意力机制的事件元素检测方法。利用事件触发词的类型信息和对应的位置信息,将BERT预训练词向量、词性向量与触发词类型向量、触发词位置向量相拼接,拼接后得到的文本向量作为Bi LSTM网络层的输入,在Bi LSTM网络层的基础上添加了注意力层,更好地去获取触发词周围的事件元素信息,最后经过softmax层输出得到事件元素的检测结果。3.针对并行完成事件类型和事件元素的检测任务,提出了一种基于联合模型的事件抽取方法。本文将BERT预训练向量、词性向量、句法分析向量和语义依存向量相结合作为神经网络层的输入。经过Bi LSTM网络层对输入的文本向量进行的文本特征提取后,将特征信息作为GCN网络层的输入,利用GCN网络层进行下一步的文本特征提取,在GCN完成相对应的计算过程后,由联合检测层的两个的分类器完成分类工作。
【Abstract】 With the rapid development of computer technology and the scale of the Internet,it is particularly important to extract useful information from the growing mass of network information and present it in the form of structured text,and information extraction technology has attracted much attention as a solution to deal with this problem.Among them,event extraction is an important research direction in the field of information extraction and one of the most challenging tasks in the field of information extractionEvent extraction is to extract the events that users care and are interested in from the massive unstructured text described in natural language,and then save and display them in a structured form to facilitate users to quickly obtain the core information of the events and grasp the corresponding progress of the events.The event extraction task can be divided into two sub-tasks from the processing flow,which are event trigger word extraction and event element extraction.There are problems in traditional event extraction methods such as easy to ignore contextual information and inadequate extraction of key features of text.In order to solve the above problems,this paper investigates the event extraction of Chinese text using deep learning methods,and the specific research work is as follows:1.A recurrent neural networks-based Chinese event trigger word detection method is proposed for the problem of Chinese event type recognition and classification.In this paper,we use the BERT model to pre-train the Chinese corpus to obtain the BERT pretrained word vectors,and use the LTP tool to analyze and process the sentence to obtain the lexical vectors,semantic dependency vectors and syntactic analysis vectors of words.The four text vectors are stitched together as the input to the Bi LSTM network layer.The information feature representation of the text is calculated by the Bi LSTM network layer,and the detection results of event-triggered words are finally obtained after processing by the CRF sequence annotation layer.2.An event element detection method combining event trigger word information and attention mechanism is proposed for the recognition and classification of Chinese event elements.Using the type information and corresponding location information of the event trigger word,the BERT pre-trained word vector and lexical vector are spliced with the trigger word type vector and trigger word location vector,and the text vector obtained after splicing is used as the input of the Bi LSTM network layer,and the attention layer is added on top of the Bi LSTM network layer to better obtain the event element information around the trigger word,the Softmax layer is used as the output to get the detection results of event elements.3.A joint model-based event extraction method is proposed for accomplishing the task of detecting event types and event elements in parallel.In this paper,syntactic analysis vectors and semantic dependency vectors are combined with BERT pretraining vectors and lexical vectors as the input of the neural network layer.After the text feature extraction of the input text vectors by the Bi LSTM network layer,the feature information is used as the input of the GCN network layer,and the next step of text feature extraction is performed using the GCN network layer,and the classification work is completed by the two trained classifiers of the joint detection layer after the GCN completes the corresponding computation process.