.png){kind=icon}
.png){kind=icon}
Return
Sentimental image generation with image quality assessment
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2026.113269.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• Distinct from previous acoustic, linguistic, and textual augmentation, our work is the first to enhance textual ABSA by generating sentimental images from scratch. The proposed SIGQA framework enables targeted visual augmentation to strengthen textual feature extraction, achieving state-of-the-art performance with low extra inference overhead. • We introduce a novel region-level no-reference IQA in SIGQA, which abandons holistic evaluation. By segmenting images for fine-grained assessment, our method provides more precise quality measurements tailored for ABSA. It aligns quality signals with aspect-related evidence and offers more reliable guidance for multimodal fusion and training. • Experiments show SIGQA is generalizable and robust, achieving state-of-the-art results in both textual and multimodal ABSA, even outperforming human-annotated images. Generated images also greatly improve cross-domain transfer, showing stronger cross-domain semantic transfer ability than text-only representations.
Keywords:
Sentimental image generation
Aspect-based sentiment analysis
Image quality assessment
Multimodal learning
Textual augmentation
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W