【作者】 罗斌；

【导师】 漆桂林；

【作者基本信息】 东南大学 ， 计算机科学与技术， 2017， 硕士

【摘要】 近年来,基于维基百科的知识库构建得到人们的大量关注,维基百科中各种不同的语义信息被挖掘出来,而其中最重要的语义信息之一就是实体类别(Type)信息。目前在维基百科中推断实体类别的数据源主要分为两种:一种是从维基百科半结构化数据,如信息框(Infobox)和文章分类(Category)中推断实体类别;另一种是在维基百科文章摘要部分进行实体类别推断。但是现有工作未考虑维基百科中大规模的文章正文,而其中恰恰隐含了大量的实体类别信息。为了进一步补充现有知识库中的实体类别信息,本文旨在从维基百科中的全部文本数据(文章摘要+正文)中进行实体类别推断。目前,基于文本的实体类别推断在自然语言处理领域存在一些相关研究,但这些方法存在如下问题:1)通过统计机器学习方法进行实体类别推断的方法需要大量训练数据,而现有获取的训练数据中存在大量噪声,2)严重依赖于人工定义的特征或者规则,3)无法说明为什么实体被推断为某个类别,即不具备可解释性。因此,研究如何在克服以上问题的基础上有效地在维基百科全部文本中进行实体类别推断是一个值得研究的课题。本文提出了一个从维基百科文本获得训练数据的方法,基于这些训练数据构造了一个图模型,从中抽取出用于实体类别推断的模式。该模式抽取方法不依赖于人工定义的特征,同时用于实体类别推断的模式具有可解释性。具体而言,本文的主要贡献有以下三点内容:1)本文提出了一种通过启发式规则与随机游走获取训练数据的方法。相比之下,目前一些从维基百科获取训练数据的方法会造成大量噪声,对实验结果产生不良影响。2)本文提出了一种基于维基百科训练数据的细粒度实体类别推断方法。本文构建一个基于词嵌入的图模型,通过Bootstrapping算法从中抽取出用于实体类别推断的模式。然后尝试将实体所在的上下文与这些模式进行匹配,如能够匹配某一模式,则句子中的实体将被推断为模式对应的类别,同时模式可作为推断得到的实体类别的解释。实验结果表明,本文方法在不同的数据集上均优于现有的基于文本的实体类别推断方法。3)本文提出的方法能够用于扩充维基百科中的实体类别信息。在针对维基百科的全部文本使用本文提出的方法进行实体类别推断后,相比于YAGO,共得10,417,582条新的类别信息,平均为每个实体扩充了 2.49个类别。更多还原

【Abstract】 Knowledge base construction from Wikipedia has attracted much attention in the last few years.So far,various semantic information has been mined from Wikipedia,and one of the most valuable semantic information is type information.Currently,there are mainly two kinds of data sources for automatically inferring entity types in the whole Wikipedia:one is the semi-structured data such as infoboxes or categories in Wikipedia;the other is the abstracts of Wikipedia articles.However,the above data sources miss large scale main text of the articles in Wikipedia.Such main text does contain plenty of type information.In order to enrich the type inormation in knowledge bases,this thesis proposes an approach to infer entity types using all textual data from Wikipedia(i.e.article abstract + main text).At present,there has been some related work for the task of text-based type inference in natural language processing,but there are following problems in their work:1)a lot of noise exists in the automatically acquired training data,2)they rely heavily on manual defined features or rules,3)they can not explain why the entity has the inferred types.Therefore,it is worth to study how to overcome the above problems and infer entity types in all textual data from Wikipedia.This thesis proposes an approach to get training data from Wikipedia,and a graph model is constructed based on these training data,patterns for type inference are extracted from the graph model.The pattern extraction approach do not rely on manual defined features or rules,meanwhile patterns for type inference can explain why the entity has the inferred types.The major work and contributions in this thesis conclude the following three aspects:1)This thesis proposes a method for automatically acquiring training data with heuristic rules and random walk.The existing methods of extracting training data from Wikipedia may contain noisy trining data,and these noisy trining data will have a negative impact on the result.2)This thesis proposes an fine-grained entity typing approach based on the high quality training data from Wikipedia.A graph model based on word embedding is constructed,and a Bootstrapping algorithm is used to extract the patterns for type inference.Then,the patterns are used to match entity context,and if one pattern can match the context of some entity,the corresponding type of the pattern will be inferred as the type of the entity.Also,the pattern can be treated as the interpretation for the results of entity type inference.The experimental results showed that the proposed approach ourperforms the existing methods for fine-grained entity type inference on different datasets.3)The approach proposed in this thesis can be used to infer entity types in Wikipedia.After performing entity type inference using the proposed approach on the whole textual data in Wikipedia,we get 10,417,582 new type information compared with YAGO,and averagely enriched 2.49 types for each entity.更多还原