Enhancing aspect-based sentiment analysis with multiple-knowledge promotion and multi-perspective noise filtering

Abstract Multimodal aspect-based sentiment analysis (MABSA) aims to integrate multimodal sentiment cues to understand users’ sentiment polarity toward the targeted aspect for supporting relevant decision-makings in various applications. Although abundant MABSA approaches have been developed, they ca...

Full description

Saved in:
Bibliographic Details
Main Authors: Juan Yang, Yali Xiao, Xu Du
Format: Article
Language:English
Published: Springer 2025-07-01
Series:Complex & Intelligent Systems
Subjects:
Online Access:https://doi.org/10.1007/s40747-025-02034-0
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Abstract Multimodal aspect-based sentiment analysis (MABSA) aims to integrate multimodal sentiment cues to understand users’ sentiment polarity toward the targeted aspect for supporting relevant decision-makings in various applications. Although abundant MABSA approaches have been developed, they cannot simultaneously utilize multiple types of knowledge and fail to effectively filter out the potential noise from different perspectives, which can reduce the discriminative ability of multimodal representations and prevent models from achieving more satisfactory performance. This study proposes an enhanced MABSA approach with multiple-knowledge promotion and multi-perspective noise filtering for addressing the research gaps. Specifically, it converts the image into an implicit sequence of token embeddings to implement fine-grained image-aspect interactions, which facilitates noise filtering from the visual semantic perspective. Meanwhile, it constructs the graph structure to fully utilize the context-aware fused semantics, syntactic dependencies, and domain sentiment knowledge, and then performs aspect-oriented image-sentence interactions to extract significant sentiment cues and discard noise from the visual-textual semantic perspective. Last but not least, it further filters out noise at the multimodal-fusion level to generate more robust and discriminative multimodal representations for enhancing aspect-based sentiment analysis performance. Finally, extensive experiments have been conducted on two popular Twitter datasets to demonstrate the superiority and effectiveness of our proposed method.
ISSN:2199-4536
2198-6053