Triple Tag Network for Aspect-Level Sentiment Classification

Abstract

The purpose of aspect-level sentiment classification is to determine the sentiment polarity of specific aspects in a sentence. Since the attention-based neural network model cannot accurately capture the connection between aspects and opinion words, many studies have adopted graph neural networks (GNN) to capture the structural information in the dependency tree of sentences, establish the relationship between aspects and opinion words. However, the sentence structure information used by these models when making sentiment polarity prediction is not comprehensive enough. Therefore, we propose a triple tag network (TTN) to imitate the operation of humans when judging the sentiment polarity. Specifically, when people determine the sentiment polarity of a certain aspect in a sentence, they generally consider the two tags of part-of-speech and dependency. On this basis, we add an additional distance tag to fit our model. We have conducted many experiments on three benchmark datasets. The experimental results show that our method can well capture the connection between aspects and opinion words, and further improve the performance of the graph attention networks (GAT).