RAMP: Retrieval Augmented Deep Perturbation Prompt Learning for Multimodal Aspect Based Sentiment Analysis

Multimodal aspect based sentiment analysis (MABSA) aims to infer the sentiment polarity toward a specific aspect by jointly leveraging textual and visual evidence. Despite recent progress, existing methods still suffer from three major limitations: they often rely on static shared prompts, have limited access to aspect-relevant external knowledge, and remain vulnerable when multimodal cues are noisy, incomplete, or weakly aligned with the target aspect. To address these issues, we propose RAMP , a unified framework that integrates aspect-aware retrieval, dynamic prompt generation, missing-cue compensation, hierarchical prompt encoding, and deep perturbation optimization. Specifically, RAMP first retrieves semantically relevant multimodal instances from an external memory bank to introduce aspect-aware contextual evidence beyond the current sample. It then converts the retrieved evidence into instance-specific prompts through a context-aware dynamic prompting mechanism, while adaptively compensating for unreliable or incomplete modalities. The generated prompts are further refined by a hierarchical prompt encoder and injected into multiple Transformer layers through deep perturbation, enabling shallow retrieved knowledge and deep semantic guidance to be consistently integrated. Finally, an aspect-guided aggregation module produces the sentiment prediction. Extensive experiments on Twitter-2015, Twitter-2017, and MASAD demonstrate that RAMP consistently outperforms strong baselines in both Accuracy and Macro-F1.