基于词向量语义分类的微博实体链接方法  

An Entity Linking Method for Microblog Based on Semantic Categorization by Word Embeddings

作　　者：冯冲[1] 石戈[1] 郭宇航[1] 龚静[1] 黄河燕[1,2] 

机构地区：[1]北京理工大学计算机学院,北京100081 [2]北京市海量语言信息处理与云计算应用工程技术研究中心,北京100081

出　　处：《自动化学报》2016年第6期915-922,共8页Acta Automatica Sinica

基　　金：国家重点基础研究发展计划(973计划)(2013CB329303);国家高技术研究发展计划(863计划)(2015AA015404);国家自然科学基金(61502035);高等学校博士学科点专项科研基金(20121101120026)资助~~

摘　　要：微博实体链接是把微博中给定的指称链接到知识库的过程,广泛应用于信息抽取、自动问答等自然语言处理任务(Natural language processing,NLP).由于微博内容简短,传统长文本实体链接的算法并不能很好地用于微博实体链接任务.以往研究大都基于实体指称及其上下文构建模型进行消歧,难以识别具有相似词汇和句法特征的候选实体.本文充分利用指称和候选实体本身所含有的语义信息,提出在词向量层面对任务进行抽象建模,并设计一种基于词向量语义分类的微博实体链接方法.首先通过神经网络训练词向量模板,然后通过实体聚类获得类别标签作为特征,再通过多分类模型预测目标实体的主题类别来完成实体消歧.在NLPCC2014公开评测数据集上的实验结果表明,本文方法的准确率和召回率均高于此前已报道的最佳结果,特别是实体链接准确率有显著提升.As a widely applied task in natural language processing（NLP）, named entity linking（NEL） is to link a given mention to an unambiguous entity in knowledge base. NEL plays an important role in information extraction and question answering. Since contents of microblog are short, traditional algorithms for long texts linking do not fit the microblog linking task well. Precious studies mostly constructed models based on mentions and its context to disambiguate entities,which are difficult to identify candidates with similar lexical and syntactic features. In this paper, we propose a novel NEL method based on semantic categorization through abstracting in terms of word embeddings, which can make full use of semantic involved in mentions and candidates. Initially, we get the word embeddings through neural network and cluster the entities as features. Then, the candidates are disambiguated through predicting the categories of entities by multiple classifiers. Lastly, we test the method on dataset of NLPCC2014, and draw the conclusion that the proposed method gets a better result than the best known work, especially on accurancy.

关 键 词：词向量 实体链接 社会媒体处理 神经网络 多分类 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术] TP393.092[自动化与计算机技术—计算机科学与技术]