Abstract
INTRODUCTION: The growing dependence on real-time social media data for situational awareness during disasters highlights the need for sophisticated sentiment analysis systems specifically designed for crisis contexts. Traditional sentiment analysis approaches, including shallow machine learning techniques and standard transformer-based models, exhibit significant limitations in addressing the linguistic complexity inherent to disaster-related discourse, including sarcasm, domain-specific lexicons, and ambiguous emotional signals, thereby restricting their applicability in high-stakes scenarios. These methods often overlook the integration of structural, symbolic, and semantic knowledge that is crucial for interpreting nuanced sentiment under crisis conditions. METHODS: This research proposes a unimodal (text-only) sentiment analysis framework incorporating a Sentiment-Enhanced Multi-Branch Network (SentEMBNet) and a Polarity-Aligned Curriculum Optimization (PACO) strategy to overcome these challenges. SentEMBNet combines lexicon-informed embeddings, graph-based syntactic modeling, and transformer-driven abstractions within a cohesive architecture to capture sentiment manifestations from granular linguistic features to broader discourse patterns. PACO employs confidence-sensitive curriculum learning, sentiment regularization, and contrastive embedding alignment to enhance adaptability across diverse semantic and structural polarity scenarios. RESULTS AND DISCUSSION: Empirical evaluations on disaster-specific and social media sentiment datasets demonstrate marked advancements in accuracy, resilience, and generalizability, particularly under conditions of noisy, unstructured inputs and domain shifts. This framework closes the semantic gap between unstructured social signals and actionable insights, strengthening the capability for sentiment-aware decision-making in disaster management systems.