A3SN: Amplifying Aspect-Sentence Awareness for Aspect-based Sentiment Analysis

ACL ARR 2025 May Submission1677 Authors

Abstract: Aspect-based sentiment analysis (ABSA) is a vital natural language processing task that extracts fine-grained sentiments for specific text aspects, yielding nuanced insights from customer reviews, social media and beyond. Native Sparse Attention (NSA), an efficient alternative to dense attention-based methods, excels at modeling long-context dependencies, local precision and fine-grained features. However, NSA faces three ABSA challenges: (1) Aspect overlap, where proximate aspects trigger selection conflicts; (2) Sparse misses, omitting critical sentiment cues in sparse selections; and (3) Global noise, where token compression dilutes aspect-specific signals. To address these challenges, we introduce a simple yet effective method, Amplifying Aspect-Sentence Awareness (A3SN), a novel method that enhances aspect-sentence interactions by doubling attention weights between aspects and contextual sentences, capturing subtle dependencies precisely. Experimental results on three benchmark datasets demonstrate A3SN's effectiveness, outperforming state-of-the-art (SOTA) baseline models while maintaining simplicity.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: Natural language processing, aspect-based sentiment analysis, sparse attention, multihead attention

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 1677