Fine-grained Sentiment Analysis based on Graph Network and Capsule Network

Hongwei Tang
Northwest Institute for Non-Ferrous Metal Research, Xi' an, 710016, China
Xueru Bai
School of Computer science and Engineering, Xi' a University of Technology, Xi' an, 710048, China
Rong Fei
School of Computer science and Engineering, Xi' a University of Technology, Xi' an, 710048, China

Abstract:

Sentiment analysis is an important research branch in the field of NLP, and its goal is to judge the sentiment polarity of a given corpus. In recent years, Graph Convolutional Networks (GCN) are widely used in fine-grained sentiment analysis tasks, and the existing research results show that GCNs are highly sensitive to syntactic information in text. However, the lack of a constraint mechanism in the GCN convolution process leads to a lot of redundant information in the network, and most of the models input the hidden layer results of GCN directly into the fully connected layer for classification, which does not fully utilize the global and syntactic information of the text. To alleviate these shortcomings, this paper proposes a model that serially fuses GCN and capsule networks. Specifically, firstly, the distance index is introduced in the embedding layer to improve the degree of association between aspect words and context-dependent words, which improves the efficiency of the downstream task processing; secondly, the distance decay function is designed as a constraint in the convolution process of the GCN, which reduces the noise in the network; lastly, the capsule network is used to replace the pooling layer in the GCN, and the dynamic routing mechanism is utilized to complete the updating of the network parameters and homologous capsule aggregation to fully exploit the feature information. In addition, the introduction of the "gate" mechanism in the routing process reduces the irrelevant parameters in the iteration and improves the aggregation efficiency, and the experimental results on five public datasets all perform well.