From Overall Sentiment to Aspect-Level Insights: A Pretraining Strategy for Unsupervised Aspect-based Sentiment Analysis
Simone Prete, Giovanni Maria Biancofiore, Fedelucio Narducci, Eugenio Di Sciascio, Tommaso Di Noia · 2025
Aspect-Based Sentiment Analysis (ABSA) aims to identify sentiments associated with specific aspects within a text.It plays a crucial role in applications such as product reviews and customer feedback analysis, where understanding nuanced opinions is essential.However, progress in ABSA remains constrained by the need for fine-grained labeled data, limiting the applicability of supervised models in real-world scenarios.In this study, we propose an unsupervised transformer-based approach that leverages sentence-level sentiment annotations to induce aspect-level sentiment representations.By supervising attention distributions during pretraining, our model learns to aggregate token-level sentiment cues into contextaware aspect sentiment predictions aligned with sentence-level supervision.We further introduce an attention-based correction mechanism to refine aspect sentiment classification by accounting for the local context of each aspect term.Evaluated on benchmark datasets including Restaurants, Laptops, and Twitter domains, our method outperforms unsupervised baselines on aspect category classification while remaining comparable with strong supervised baselines on aspect term sentiment tasks.These results demonstrate that attention-guided pretraining enables robust, domain-adaptive ABSA without requiring aspect-level supervision.