融合图卷积神经网络的文本情感分类

文档级别情感分类旨在预测用户对评论文本的情感极性标签。最近研究发现,利用用户和产品信息能有效地提升情感分类性能,然而,现有大多数研究只关注用户与评论、产品与评论的信息,忽略了用户与用户、产品与产品之间的内在关联,因此,本文提出一种融合图卷积神经网络的文本情感分类模型。首先,根据数据集构建了用户与用户关系图、用户与产品关系图;然后,融合两种关系图形成异质图,并使用图卷积神经网络学习用户与用户、产品与产品之间的内在联系,获得更好的用户和产品表示;最后,使用融合CNN的用户注意力和产品注意力机制的分层网络进行情感分类。实验结果表明,在公开数据集IMDB、Yelp2013和Yelp2014上,本文提出的模型能取得较好的分类效果。

Sentiment classification combining graph convolution neural network

Document-level sentiment classification aims to predict users sentiment polarity labels on review text. Recent studies have found that incorporating user and product information can effectively improve the performance of review sentiment classification. However, most of the pervious arts only concentrate on the relation between users and reviews, products and reviews, ignoring the inhere correlation between users and users, products and products. Therefore, we propose a sentiment classification combining graph convolution neural network. Firstly, we construct a relation graph of user-to-user and a relation graph of user-to-product based on the datasets and merge them into a heterogeneous graph. Secondly, utilize the graph convolutional neural network to learn the inhere correlation between users and users, products and products to obtain better user and product representation. Finally, use a hierarchical user attention and product attention network combined with CNN to make sentiment polarity classification. Experiments on benchmark datasets of IMDB, Yelp2013 and Yelp2014 show that our method can achieve state-of-the-art performance.