An aspect-centric, noise-resilient cross-modal dynamic fusion framework for fine-grained sentiment analysis

Abstract

Multi-modal Aspect-Based Sentiment Analysis (MABSA) aims to identify sentiment polarity toward specific aspects in text-image pairs. Current models often fail to achieve precise cross-modal alignment and struggle with visual background noise in real-world data, leading to inaccurate predictions. To address these, we propose an aspect-centric, noise-resilient cross-modal dynamic fusion (ANDF) framework. Our framework introduces three key innovations. First, an Aspect-Prompt Encoding strategy generates aspect-aware query features for precise, fine-grained alignment of text-visual features. Second, an Aspect-Centric Visual Disentanglement module uses dynamic gating to effectively decouple foreground and noisy background visual information, generating robust aspect-aware prototypes. Third, a Prototype-Guided Fusion module integrates multi-source features via a Transformer-based structure, enabling dynamic and complementary feature aggregation. Extensive experiments on the Twitter-2015 (2166 samples) and Twitter-2017 (5818 samples) datasets demonstrate the effectiveness of our approach. On Twitter-2015, our model achieves the highest Accuracy among all compared baselines (79.27%), outperforming the previous best method AMIFN (78.69%) by 0.58 percentage points. On Twitter-2017, ANDF attains a leading F1-score of 71.76%, surpassing all baselines on this more challenging benchmark. Controlled perturbation experiments further show the resilience of ANDF under partial occlusion, fine-grained noise, semantic mismatch, and structural disruption. Ablation studies indicate that our core modules effectively enhance cross-modal semantic alignment and noise suppression.