Aspect-level Sentiment Classification Method Based on Graph Convolutional Network and Interactive Attention Mechanism

Abstract

Aspect-level sentiment classification often uses syntactic information such as dependency trees to improve model performance, which makes aspect words far away from emotional words; When the text is modeled as a graph model, the interactive information between context and aspect words is ignored in the graph convolutional network, resulting in insufficient semantic information and local feature information. To solve these two problems, an aspect-level sentiment classification method based on graph convolutional network and interactive attention mechanism is proposed. By fine-tuning RoBERTa, the model strengthens the connection between aspect words and emotional words, shortening the distance. The graph convolutional network constructs the induced tree with syntactic dependence between framework words and aspect words into dependency tree. Finally, the interactive attention mechanism is used to study the interactive information between context and aspect words. The experimental results show that contrast to the baseline model, this model improves the accuracy and F1 value.