Dependency-Aware Graph Transformer with Hierarchical Syntax Encoding for Aspect-Level Sentiment Classification
Published:
Please cite:
@inproceedings{shaokun_ewsa2026,
title={Dependency-Aware Graph Transformer with Hierarchical Syntax Encoding for Aspect-Level Sentiment Classification},
author={Shaokun Liu, Mieradilijiang Maimaiti, Nuermaimaiti Youliwasi, Nilufar Abdurakhmonova, Wu Le, Zhoufei Xie, Jiawei Chen, and Wushour Silamu},
journal={Expert Systems with Applications (ESWA)},
year={2026},
}
Abstract
Aspect-Level Sentiment Classification (ALSC) aims to identify the sentiment polarity of specific aspects within textual content. Although graph neural networks (GNNs) that utilize dependency syntax trees have achieved strong performance in ALSC tasks, existing approaches commonly struggle to distinguish the significance of different dependency relations and fail to sufficiently capture semantic interactions, thus restricting their ability to detect implicit sentiments. To tackle these challenges, we introduce the Dual-stage Dependency Enhanced Graph Transformer (DSDGT), a sequential architecture integrating the Transformer and Graph Convolutional Network (GCN) modules with an enhanced dependency mechanism. The Transformer component is designed to capture the Global information, while the GCN models local dependency structures. A two-phase feature processing strategy is adopted to merge both original and enhanced dependency features. Experimental results on multiple benchmark datasets demonstrate that DSDGT outperforms state-of-the-art methods in terms of accuracy and macro-F1 scores, validating its superior performance.
[PDF]