基于类型间关系学习的细粒度实体分类

Type relation learning for fine grained entity type classification

细粒度实体分类旨在为构建知识图谱过程中所抽取的实体或实体提及确定一个或多个层次化、细粒度的类型,以便更好地为下游任务提供支持。现有细粒度实体分类方法存在细粒度分类精度不高、部分实体难以有效分类的问题。另一方面,直观来说,掌握细粒度类型之间的语义区别有助于实体的细粒度分类。但由于已有面向该任务的数据集缺少可用于学习细粒度类型间语义差别的数据,因此目前没有将细粒度类型之间的语义区别应用于细粒度实体分类的研究。为此,本文提出一种基于Freebase知识库学习细粒度类型语义区别的方法,并将学习到的语义信息应用在细粒度实体分类任务中。具体地,利用SPARQL从Freebase中获取类型之间的关系数据,据此学习细粒度实体类型之间的语义区别信息,进而结合实体提及及其上下文的文本信息进行细粒度实体分类。实验表明,本文提出的方法可以有效学习细粒度类型之间的语义区别,能够达到提升细粒度实体分类准确率的效果。

In order to support downstream tasks better,the aim of fine grained entity type classification is to assign one or more hierarchical and fine grained types to the entity or mention extracted in the processing of knowledge construction.Prior work has the problems of lower accuracy or being unable to assign the right types.Intuitively,the semantic distinction is useful of fine grained entity typing.There is no research on this aspect since the datasets of fine grained entity type contain no data for learning the semantic distinction between types.A method which learns the semantic distinction of fine grained types using the knowledge base Freebase is proposed,and the learned semantic distinction is used in the task of fine grained entity type classification.Specifically,SPARQL is used to acquire the relations data from Freebase,and the semantic distinction is learned from those acquired data,then the fine grained type is assigned to entity according to the learned semantic distinction of types together with the semantic information learning from mention and context.Experimental results and analysis on datasets demonstrate that the proposed model can learn the semantic distinction of types,and the results of the proposed model outperforms other state-of-the-art methods.