【作者】 蒋凯；

【导师】 江大伟；

【作者基本信息】 浙江大学 ， 工程硕士（专业学位）， 2022， 硕士

【摘要】 随着互联网的发展以及大数据时代的到来,每天都会有大量的数据产生,这其中就有很多文本数据。个人已经很难从大量的数据中找到有用信息,必须借助于机器的前期处理。然而便于人类理解的非结构化自然语言文本却不便于机器直接进行计算处理。属性抽取作为信息抽取技术中的一种,研究从自然语言文本中提取实体属性的结构化信息,得到了广泛的研究。属性抽取是许多自然语言处理任务的基础工作。本文主要研究从给定中文文本中抽取给定实体的属性,提出的孪生BERT模型以提高在样本量不充足情况下模型的抽取性能,主要工作如下:1.本文提出了基于问答模式的属性值抽取模型。模型首先提出与文本对应的属性关系和实体的问题,然后利用问答系统在给定的文本中抽取出问题的答案,抽取的答案即为实体对应的属性值。基于问答模式的属性值抽取使得模型更偏向了对文本的理解,从文本以及实体和属性关系信息中提取的特征具有更强的泛化性。本文使用指针网络从提取的特征中解码出属性值,区别于传统的利用序列标注模型来提取属性值。模型中使用预训练模型BERT编码文本信息。BERT模型是基于大量未标注数据训练的语言模型,能够有效提取文本中的词法信息,为属性值的解码提供良好的文本特征。2.本文提出了基于孪生BERT的属性关系判别模型,其中也是利用BERT预训练模型提取文本特征信息。BERT模型在提取文本特征时能有效地提取文本的词法信息,但对于句法信息的提取尚有不足。针对BERT模型提取文章句法信息不足的问题,本文将孪生网络结构应用于BERT模型,提出孪生BERT模型,用于改良模型对句法信息的提取。孪生网络模型的训练需要比较样本间的相似性,本文用样本属性关系的相似来代表样本的相似性。3.本文将属性抽取任务分为属性关系判别与属性值抽取两个子任务,首先利用孪生BERT模型判别文本中存在的属性关系,得到实体和属性关系对应的问题,再通过基于问答模式的属性值抽取模型提取文本中的属性值。本文在百度百科数据集和Du IE2.0数据集上验证了基于问答模式的属性值抽取模型与基于孪生BERT的属性关系判别模型的有效性,以及将属性抽取任务分为属性关系判别与属性值抽取的两个子任务的可行性。更多还原

【Abstract】 At the Internet era,a lot of text data is generated every day.It is already difficult for individuals to find useful information from a large amount of data,and must resort to pre-processing by machines.However,the unstructured natural language text data that is easy for humans to understand is not easy for machine calculation and processing.Information extraction technology is the study of converting unstructured text data into structured data that is convenient for machine calculation and processing.This article focuses on extracting the attributes of a given entity from a given text.The main work of this paper is as follows:1.This paper proposes an attribute extract model based on the question and answer mode,which proposes the attribute relationship and entity questions corresponding to the text.The answer to the question is extracted from the given text through the question and answer system,and the extracted answer is the attribute value corresponding to the entity.The attribute value extraction based on the question and answer mode makes the model more inclined to the understanding of the text,and the features extracted from the text and the entity and attribute relationship information have stronger generalization.This paper uses the pointer network to decode the attribute value from the extracted features,which is different from the traditional sequence labeling model to extract the attribute value.The pre-trained model BERT is used in the model to encode text information.The BERT model is a language model trained on a large amount of unlabeled data,which can effectively extract lexical information in the text and provide good text features for the decoding of attribute values.2.This paper proposes an attribute relationship discriminant model based on twin BERT.The BERT model can effectively extract the lexical information of the text when extracting text features,but it is still insufficient for the extraction of syntactic informa-tion.Aiming at the problem of insufficient syntactic information extracted by the BERT model,this paper applies the twin network structure to the BERT model,and proposes the twin BERT model to improve the model’s extraction of syntactic information.The training of the twin network model needs to compare the similarity between samples.This paper uses the similarity of the sample attribute relationships to represent the simi-larity of the samples.3.In this paper,the attribute extraction task is divided into two subtasks: attribute relationship discrimination and attribute value extraction.First,the twin BERT model is used to discriminate the attribute relationship existing in the text,and the corresponding question of the attribute relationship is obtained,and then the attribute value extraction model based on the question and answer mode is used.Extract the attribute value in the text.This paper verifies the effectiveness of the attribute value extraction model based on the question and answer mode and the attribute relationship discrimination model based on twin BERT on the Baidu Encyclopedia dataset and the Du IE2.0 dataset,and divides the attribute extraction task into attribute relationship discrimination and attribute value extraction.The feasibility of the two tasks.更多还原