MAMS-Enriched: Structural and Graph-Based Representations for Aspect-Based Sentiment Analysis

Abstract

Aspect-Based Sentiment Analysis (ABSA) re-mains challenging in sentences containing multiple aspects with conflicting sentiment polarities, where models must correctly associate opinion expressions with their corresponding targets. Although recent ABSA research has predominantly focused on increasingly complex neural architectures, comparatively limited attention has been directed toward the structural quality and linguistic richness of dataset representations. Ex-isting benchmark datasets, including MAMS, provide aspect-level sentiment annotations but lack explicit syntactic and relational information required for structure-aware learning and interpretable graph-based modeling. This paper introduces MAMS-Enriched, a syntax-enhanced and graph-oriented ex-tension of the MAMS dataset for Aspect-Based Sentiment Analysis. The proposed resource augments each sentence with token-level Part-of-Speech (POS) tags, dependency relations, aspect-relative distance features, lexical sentiment priors, and aspect-indicator representations generated through an auto-mated enrichment pipeline integrating spaCy, VADER, and BERT-based token alignment. Each aspect term is transformed into an independent graph instance with bidirectional depen-dency connectivity, enabling direct compatibility with graph-learning frameworks such as PyTorch Geometric. To evaluate the effectiveness of the enriched resource, experiments were conducted using lightweight BiLSTM and Graph Convolutional Network (GCN) models under enriched and ablation settings. Experimental results demonstrate that the enriched dataset improves representation quality and predictive performance. In particular, BiLSTM accuracy improved from 47.18% to 67.12%, corresponding to an absolute improvement of 19.94 percentage points after enrichment. Ablation analysis further confirms the importance of dependency structure, where removal of dependency relations reduced GCN accuracy from 52.25% to 46.41%. Multi-seed evaluation additionally demonstrates con-sistent performance stability and reproducibility across random initializations. The findings demonstrate that direct structural enrichment significantly improves dataset quality for aspect-level sentiment analysis while enhancing interpretability, graph compatibility, and experimental consistency. The proposed resource provides a reusable foundation for future research in syntax-aware and graph-based ABSA systems.