Relevance-aware visual entity filter network for multimodal aspect-based sentiment analysis

Multimodal aspect-based sentiment analysis, which aims to identify the sentiment polarities over each aspect mentioned in an image-text pair, has sparked considerable research interest in the field of multimodal analysis. Despite existing approaches have shown remarkable results in incorporating external knowledge to enhance visual entity information, they still suffer from two problems: (1) the image-aspect global relevance. (2) the entity-aspect local alignment. To tackle these issues, we propose a Relevance-Aware Visual Entity Filter Network (REF) for MABSA. Specifically, we utilize the nouns of ANPs extracted from the given image as bridges to facilitate cross-modal feature alignment. Moreover, we introduce an additional “UNRELATED” marker word and utilize Contrastive Content Re-sourcing (CCR) and Contrastive Content Swapping (CCS) constraints to obtain accurate attention weight to identify image-aspect relevance for dynamically controlling the contribution of visual information. We further adopt the accurate reversed attention weight distributions to selectively filter out aspect-unrelated visual entities for better entity-aspect alignment. Comprehensive experimental results demonstrate the consistent superiority of our REF model over state-of-the-art approaches on the Twitter-2015 and Twitter-2017 datasets.