摘要
为了解决传统的深度学习模型会忽略语料库中全局词共现信息所包含的非连续和长距离语义的问题。本文提出记忆图卷积神经网络(MGCNN)引入注意力机制的商品评论情感分析方法。首先提取词与词、词与文档之间的关系,以全部的词和文档作为节点,将整个数据集构造成一个异构文本图。再基于图卷积网络(GCN)来构建用于图结构数据的神经网络,利用长短期记忆网络(LSTM)提取上下文相关特征,并使用注意力层获取重要特征。多组对比实验结果表明,本方法的分类效果更好,且随着训练集数据所占比例的降低,其优势更加显著。
To solve the problem that the traditional deep learning models ignore the discontinuous and long-distance semantics existing in the global word co-occurrence information in the corpus.This paper proposes a sentiment analysis method for product reviews by introducing the attention model into the memory graph convolutional neural network.We constructed a heterogeneous text graph from a data set by taking words and documents as nodes and considering relationships among them.Then,graph convolutional network was used to extract features of graph structure data.Additionally,the long short-term memory network was employed to extract context-related features.After that,the attention layer was used to focus on important features.The results of multiple comparison experiments show that our method has a better classification effect,and as the proportion of data in the training set decreases,its advantages become more significant.
作者
许犇
徐国庆
程志宇
罗京
XU Ben;XU Guoqing;CHENG Zhiyu;LUO Jing(School of Computer Science&Engineering,Wuhan Institute of Technology,Wuhan 430205,China)
出处
《武汉工程大学学报》
CAS
2020年第5期585-590,共6页
Journal of Wuhan Institute of Technology
关键词
图卷积网络
长短期记忆网络
注意力模型
商品评论
情感分析
graph convolutional network
long short-term memory network
attention model
product reviews
sentiment analysis
