Instruction-tuned ABSA with auxiliary sentences and knowledge-enhanced graphs for implicit aspect detection
Kanwal Ahmed, Muhammad Imran Nadeem, Guanghui Wang, Fang Zuo, Zhijie Han
Abstract
Aspect-Based Sentiment Analysis (ABSA) is a critical task in Natural Language Processing (NLP), often challenged by implicit aspect expressions, ambiguous context dependencies, and limited domain-specific labeled data. Existing methods, while leveraging pre-trained language models (PLMs), struggle to effectively incorporate domain knowledge, resolve contextual ambiguities, and refine sentiment features. This paper proposes the Auxiliary Sentence-Enhanced Instruction Tuning-based Graph Convolutional Framework (ASE-ITGF), which introduces several novel elements. The auxiliary sentence generation mechanism guides RoBERTa in learning aspect-specific representations, addressing implicit aspect terms. The framework further integrates graph convolutional networks (GCNs) enriched with SenticNet knowledge for robust syntactic-semantic alignment. Finally, a bi-layer sentiment representation module enhances feature extraction, ensuring a comprehensive understanding of sentiment-context relationships. The framework achieves state-of-the-art results on benchmark datasets, including 91.23% accuracy and 86.48% F1-score on Restaurant14, 90.22% accuracy and 81.34% F1-score on Restaurant15, and 95.52% accuracy and 81.82% F1-score on Restaurant16. On Laptop and Twitter datasets, it attains 85.32% accuracy, 82.87% F1-score and 79.82% accuracy, 77.95% F1-score, respectively, significantly outperforming existing baselines. These results underscore ASE-ITGF’s effectiveness and its innovative contributions to advancing ABSA.