【作者】 刘波；

【导师】 陈阳； 姬东溟；

【作者基本信息】 东南大学 ， 电子与通信工程（专业学位）， 2019， 硕士

【摘要】 在这个信息爆炸的时代,网上蕴含着海量的有价值的企业信息,但是这些信息大多分散在不同的网站,这就导致这些数据缺乏层次性和逻辑性,不同网站的数据集之间很难实现数据的自动关联,所以对信息的智能化和规范化处理显得尤为重要。本文研究的实体链接技术主要是解决实体的多样性、歧义性、缺失性三方面的问题。根据有监督实体链接算法的三个步骤:命名实体识别、候选实体的生成和候选实体消歧,设计了一个实体链接系统,并利用该系统成功构建了一个企业领域的知识图谱。论文的具体工作可以归纳如下:(1)选择维基百科中文版、百度百科和互动百科作为背景知识库来构建多源知识库,使用基于Att-BiLSTM-CRF中文命名实体识别模型来获得实体指称,提出了一种结合上下文匹配策略和知识库信息检索策略的实体指称扩展方法,最后生成了一个具备高召回率和高准确率的候选实体集合。(2)提出了两种融合神经网络和余弦相似度的候选实体排序算法,提出了空实体判定方法。设计不同场景对比实验,结果表明,选择融合CNN和余弦相似度的候选实体排序算法,并添加空实体判定方法得到的候选实体消歧算法效果最优。(3)结合上述候选实体生成算法和候选实体排序算法作为本文的实体链接算法,设计出一个应用于企业领域的实体链接系统,并将该系统应用到构建知识图谱的过程中,使用Neo4j成功构建企业领域的知识图谱。更多还原

【Abstract】 In this era of information explosion,there is a huge amount of valuable enterprise information on the Internet.However,most of the information is scattered on different websites,which leads to the lack of hierarchy and logic.It is difficult to achieve automatic association of data from different websites.Therefore,it is particularly important to deal with information intelligently and normatively.The entity linking technology is mainly to solve the problems of diversity,ambiguity and lack of entities.According to the three steps of the supervised entity linking algorithm: named entity recognition,candidate entity generation and candidate entity disambiguation,an entity linking system is designed and used to build an enterprise knowledge graph successfully.The specific work of the paper can be summarized as follows:(1)Build a multi-source knowledge base by selecting Wikipedia Chinese,Baidu Encyclopedia and Interactive Encyclopedia as background knowledge bases.Use the AttBiLSTM-CRF Chinese named entity recognition model to obtain entity referrals.An entity referential extension method combining context matching strategy and knowledge base information retrieval strategy is proposed.Finally,a set of candidate entities with high recall rate and accuracy rate is generated.(2)Two candidate entity ranking algorithms combining neural network and cosine similarity and an empty entity decision method are proposed.The comparison experiments of different scenes are designed.The results show that the obtained candidate entity disambiguation algorithm which selects candidate entity ranking algorithm combining CNN and cosine similarity and adds the empty entity judgment method is the best.(3)The above candidate entity generation algorithm and candidate entity ranking algorithm are redefined as the entity link algorithm herein.Design an entity linking system applied to the enterprise field,and then apply this system to the process of constructing the knowledge graph.Use Neo4 j to build an enterprise knowledge graph successfully.更多还原