Supervised attention method with co-reference resolution for aspect extraction.

Aspect-based sentiment analysis (ABSA) of review texts remains a challenging problem due to the contextual interdependence of aspect evaluations across multiple sentences. Most existing approaches perform aspect extraction in isolation, limiting their ability to capture aligned semantics and multi-word (M-W) aspect terms in complex, multi-sentence (M-S) reviews. Furthermore, baseline and recent methods struggle to model long-range contextual dependencies, resulting in incomplete or fragmented aspect identifications. To address these limitations, this paper proposes a supervised hierarchical attention-based model for enhanced aspect term detection. The proposed model incorporates co-referencing resolution (CRR)-based sentence-alignment mechanism before ABSA, which resolves pronouns and referring expressions across sentences, generating semantically aligned inputs for the hierarchical attention network (HAN). This preprocessing step enhances contextual coherence, enabling the word-level and sentence-level attention mechanisms in HAN to more effectively capture aspect-related information across multi-sentence reviews. Furthermore, contextual representations are learned using word vectors enriched with semantic context, which are then processed through HAN using sequence-tagged labeled data. Experiments conducted on the SemEval-16 benchmark datasets demonstrate that the proposed model consistently outperforms strong baselines and state-of-the-art methods across the Laptop and Restaurant domains. The experimental evaluation shows that the proposed model improves F-Score by + 2.77% and + 3.52% over AspectGCN, and by + 4.16% and + 2.32% over IDGNN + BERT on the laptop and restaurant review datasets, respectively. It highlights that the inclusion of the CRR module yields an additional 3-4% gain over the HAN baseline, strengthening the effectiveness of the context-aware and alignment-driven modeling for aspect term extraction (ATE) from M-S reviews.