融合用户信息和评价对象信息的文本情感分类
李俊杰1,2,宗成庆1,2,3*

(1.中国科学院自动化研究所,模式识别国家重点实验室 北京 100190; 2.中国科学院大学计算机与控制学院,北京 100190; 3.中国科学院脑科学与智能技术卓越创新中心,北京 100190)

情感分类; 用户信息; 深度学习

Document-level Sentiment Classification Considering User Information and Aspect Information
LI Junjie1,2,ZONG Chengqing1,2,3*

(1.National Laboratory of Pattern Recognition,Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China; 2.School of Computer and Control Engineering,University of Chinese Academy of Sciences,Beijing 100190,China; 3.Center for Excellence in

sentiment classification; user information; deep learning

DOI: 10.1016/j.neucom.2017.01.121.

备注

文档级别情感分类的目的 在于预测用户对评论文本的情感倾向.目前大部分工作只关注于文档的内容而忽视了用户信息和评价对象信息.事实上,不同的用户在表达情感时选词存在着差异,并且对同一产品不同属性的关注度也会有所不同; 不同的词汇在描述不同的评价对象时,也会有着不同的情感倾向性.为了能同时考虑用户和评价对象,提出了一个基于用户和评价对象的层次化注意力网络(hierarchical user aspect attention networks,HUAAN)模型.该模型首先用一个层次化的结构编码各类信息(包括词汇、句子、评价对象、文档),然后引入基于用户和评价对象的注意力机制来建模这两类信息.为了验证HUAAN模型的有效性,在两个真实的数据集上进行实验,结果表明在融入这两类信息之后,HUAAN在同等条件下比NSC+UPA系统的准确率高.

Document-level sentiment classification aims to infer user's sentiment polarity in a review.However,most of existing methods only focus on text information and ignore user information and aspect information.Different users may use different words to express their opinions and pay their attentions to different aspects about a product.Words describing different aspects may induce different sentiment polarities.These two kinds of information are helpful to sentiment classification.To consider these two kinds of information,we propose a model called hierarchical user aspect attention networks(HUAAN),which can encode different kinds of information(word,sentence,aspect,document)in a hierarchical structure and import the user-and-aspect-attention mechanism to model user information and aspect information.Empirical results on two real-world document-level review datasets show that our model obtains the best classification in the same condition.The accuracy rate of sentiment classification is higher than the system of NSC+UPA-pro.