基于关系指示词库的开放式实体关系抽取算法  

Open Entity Relation Extraction Based on Library of Relation Word

作　　者：王月[1,2] 周刚[1,2] 南煜[1,2] 郑梓圣 田菲[1] 

机构地区：[1]信息工程大学,河南郑州450001 [2]数学工程与先进计算国家重点实验室,河南郑州450001

出　　处：《信息工程大学学报》2017年第2期242-247,252,共7页Journal of Information Engineering University

摘　　要：为解决传统实体关系抽取方法适应性较差的问题,提出基于关系指示词库的开放式实体关系抽取方法(WCORE)。根据上下文环境自动构建与人工补充的方式构建关系指示词库,两种方式相互补充,主要论述3个问题:关系三元组的确立;关系指示词信息增益的计算与关系指示词库的构建;基于关系指示词库的分类实体关系信息抽取。以真实微博文本作为语料,平均F值达到了75.90%,实验结果表明该方法具有较好的可行性和适应性。Open Relation Extraction （ORE） can eschew hand-labeled training examples, tackle an unbounded number of entity relations, and scale linearly to handle large-scale web corpus. In this paper, we propose a model to extract entity relations on the basis of library of relation words （WCORE）. After extracting relation triples, automatic relation word identification is one way to solve ORE problems. Thus, automatic construction and artificial supplement are combined to complement their results. This paper mainly discusses three issues： relation triples extraction; importance calculation of relation word and library building of relation word ; automatic identification of relation word by using the library. Evaluations over Weibo text corpus give an average F-score of 75.90% , and the resnlts indicate that this method could be open and adaptive.

关 键 词：关系三元组 开放式实体关系抽取 关系指示词 指导库 语义相似度 

分 类 号：TN929.5[电子电信—通信与信息系统]