An ANPs-Enhanced Framework Based on Cross-Modal Translation for Multimodal Aspect-Based Sentiment Analysis

Multimodal aspect-based sentiment analysis (MABSA) is an emerging task in natural language processing that focuses on the fine-grained analysis of emotional tendencies for each specific aspect within both text and associated images. Existing research primarily concentrates on utilizing different pretrained models to extract single-modal features followed by attention-based fusion. However, due to the representational differences between modalities, semantic inconsistencies often arise during modality interactions. In this paper, we propose a novel framework based on cross-modal translation enhanced with adjective-noun pairs (ANPs), termed AEFCMT. This method converts images into descriptive text, allowing for the processing of multimodal data within a unified semantic space. Additionally, we design a multi-interaction attention module that ensures the emotional cues from images are fully considered. Experimental results demonstrate that our proposed method consistently outperforms strong baseline models on two public datasets.