Multi relational dual attention graph transformer for fine grained sentiment analysis

Abstract Aspect-Based Sentiment Analysis requires precise identification of sentiment polarity toward specific aspects, demanding robust modeling of syntactic, semantic, and discourse-level dependencies. Current graph-based approaches inadequately address the complex interplay between multiple relation types and lack effective attention regularization mechanisms for interpretability. We propose the Multi-Relational Dual-Attention Graph Transformer (MRDAGT), a novel framework unifying syntactic, semantic, and discourse relations within a coherent graph architecture. Our dual-attention mechanism strategically balances local token-level interactions with aspect-oriented contextual focus while attention regularization combining entropy-based penalties and L1 sparsity constraints ensures interpretable, focused predictions. MRDAGT establishes new state-of-the-art benchmarks across multiple datasets, delivering substantial performance improvements while maintaining transparent, linguistically grounded decision-making processes essential for real-world deployment.