FITE-GAT: Enhancing aspect-level sentiment classification with FT-RoBERTa induced trees and graph attention network

Abstract

Aspect-level sentiment classification (ASC) aims to predict the sentiment polarity of aspects in a given sentence. Combining syntactic structures with graph neural networks (GNN) has been proven effective in ASC. However, this approach is susceptible to syntactic parsing errors, which may result in erroneous association between aspect words and unrelated terms. In this paper, we propose a novel approach called FT-RoBERTa induced trees Enhanced Graph Attention Network (FITE-GAT), which establishes dependencies between aspect words and sentiment words to reduce the impact of syntactic parsing errors. FITE-GAT leverages the Perturbed Masking method to generate FT-RoBERTa induced trees, guiding the connection between aspect words and sentiment words. Additionally, FITE-GAT utilizes a Graph Attention Network (GAT) under the syntactic constraints of induced trees to assign different weights to dependent edges, preventing irrelevant words from propagating to target aspects. Furthermore, non-aspect masking is employed to obtain sequences that exclusively contain aspect words, eliminating the influence of irrelevant information. Our model effectively learns sentiment information, resulting in improved performance across three different domains. Compared to TDTE-GAT, FITE-GAT improves F1-score by 2.07%, 2.53%, and 1.91% on Laptop, Restaurant, and Twitter datasets. Compared to ASGCN, FITE-GAT enhances F1-score by 2.56%, 2.27%, and 2.66% on the same datasets.