CGRN: Conflict-Aware Geometric Routing Network for Multimodal Sentiment Analysis

Multimodal sentiment analysis over text-image pairsis challenged by cross-modal conflict, where textual and visualcues express contradictory sentiment (e.g., sarcasm andirony). Most existing systems apply uniform fusion regardlessof modality agreement, leading to suboptimal performance ondifficult, contradictory samples. We present Conflict-Aware GeometricRouting Network (CGRN), a modular architecture thatcomputes a differentiable Geometric Dissonance Score (GDS)and conditionally routes each sample to either a normal fusionbranch (for low dissonance) or a conflict-specialized branch withcross-modal attention and optional sarcasm auxiliary supervision(for high dissonance). The routing threshold τ is learnable andtrained with margin-based hinge separation on conflict vs. nonconflictGDS distributions. Routing is soft and differentiableduring training, allowing gradient flow through the controller,and hard at inference for efficiency and interpretability. CGRNadditionally generates structured per-inference Conflict Reportsgrounded in geometric and routing internals, providing explicitmechanism-level traceability. On the challenging MVSA-Multipledataset containing 19,600 text-image posts, CGRN reaches 61.8%accuracy and 0.552 macro-F1, with 78.1% of conflict samplescorrectly routed to the conflict branch, achieving a 0.483 conflictsubsetF1 score. Comprehensive ablation studies demonstrate thatboth magnitude and directional divergence are critical for reliable conflict identification. The RoBERTa-based variant has 126.9Mparameters and maintains a lightweight inference profile. Codeand artifacts are made available to promote further research ingeometrically constrained multimodal routing.Index Terms—multimodal sentiment analysis, cross-modal conflict,geometric dissonance, conditional routing, sarcasm detection,explainability, transformer networks