在线用户评论细粒度属性抽取

发布时间：2018-01-15 12:30

  本文关键词：在线用户评论细粒度属性抽取　出处：《情报学报》2017年05期 　论文类型：期刊论文

  更多相关文章： 属性抽取 属性聚类 深度学习 近邻传播聚类 细粒度属性

【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。
[Abstract]:With the rapid growth of online review information and the continuous expansion of applications, the research of comment mining has been continuously concerned by the academic community. The current task of comment mining requires more and more comprehensive attributes, fine-grained and so on. Most of the existing research methods mainly focus on the extraction of the main attributes of the evaluation object. As far as possible, we can find all the user concerned attributes of the evaluation object, and express the attributes in a fine-grained manner. In this paper, a fine-grained attribute extraction method is proposed to extract product attributes comprehensively and quickly. Firstly, candidate attribute words are constructed by using high-frequency nouns. Then, the candidate attribute word vector is constructed by in-depth learning, and the candidate attribute word set is obtained by clustering the candidate attribute. Finally, the candidate attribute word set is filtered by noise, and the fine-grained product attribute set is obtained. The experimental results in the review corpus of diet, mobile phone and book show that compared with the method based on seed words. Based on the combination of artificial LDA method and affective word based method, this method can find evaluation object attributes more comprehensively, and can give fine grained attributes.
【作者单位】： 南京理工大学信息管理系;福建省信息处理与智能控制重点实验室(闽江学院);江苏省数据工程与知识服务重点实验室(南京大学);
【基金】：国家社会科学基金项目“在线社交网络中基于用户的知识组织模式研究”(No.14BTQ033) 福建省信息处理与智能控制重点实验室(闽江学院)开放课题
【分类号】：G254
【正文快照】： 1引言当前,电商平台以及社交媒体上存在大量的用户评论信息。如何从纷繁复杂的在线评论中高效挖掘用户感兴趣的信息,是社会媒体计算领域关心的重要研究问题。属性抽取作为评论挖掘研究中重要任务之一,已经引起众多学者的重视[1_2]。在正确识别用户关注属性的同时,对抽取出的属，

本文编号：1428401