在线用户评论细粒度属性抽取

摘要

随着在线评论信息数量的快速增长与应用的不断扩展,评论挖掘研究得到学术界的持续关注.当前的评论挖掘任务对属性的全面性、细粒度等要求越来越高,而多数现有研究方法主要关注评价对象主要属性的抽取.尽可能地发现评价对象的全部用户关注属性、并以细粒度方式表述属性,是一项有意义的工作.本文提出一种细粒度属性抽取方法,旨在全面、快速地抽取产品属性.本文首先利用高频名词构建候选属性词;然后通过深度学习构建候选属性词向量,在此基础上完成候选属性的聚类,得到聚类后的候选属性词集;最后对候选属性词集进行噪音过滤,得到细粒度产品属性集.在饮食、手机、图书等三个领域评论语料上的实验结果表明,相对于基于种子词的方法、基于结合人工的LDA方法及基于情感词的方法,本文方法能够更加全面地发现评价对象属性,并且能够给出细粒度的属性.%With the rapid development of online reviews and related applications,review mining has been given sustained attention in academia.Currently,aspect mining has higher requirements for comprehensiveness and fine granularity.However,most existing methods focus on mining essential product aspects.Locating all aspects concerned by users and describing aspects in a fine-grained manner is a meaningful work.Hence,in this paper,we propose a fine-grained aspect extraction method,which attempts to extract product aspects comprehensively and effectively.Specifically,we first extract candidate aspects based on frequent nouns,and then,using deep learning,we construct candidate aspect vectors for clustering synonymous aspects.Finally,we obtain aspect sets by filtering the noise in candidate aspect sets.Experimental results on a corpus of dietary,mobile phones,and books,show that,compared with the seed words-based method,LDA-based method,and sentiment words-based method,our method can comprehensively extract opinion target aspects and identify more fine-grained aspects.