A Text-Image Prompt Template-Based Model for Multimodal Fine-Grained Sentiment Analysis

This paper introduces MTGLWS, a multimodal fine-grained sentiment analysis framework that reformulates aspect-level polarity classification as a masked-token generation task. Beginning with template-based prompts containing learnable placeholders, the approach first optimizes textual prompts via T5 fine-tuning, then augments them with image-derived naturallanguage captions encoded by a CLIP captioner, thereby unifying textual and visual cues within a single prompt. Combined prompt tokens and regional visual features extracted by Faster-RCNN are fused through a self-attention stack and injected into UNITER, whose contextualized representations enable precise sentiment inference for each aspect term. To map predicted tokens back to polarity labels, the method constructs class-specific label-word sets expanded from external sentiment lexicons and filtered by cosine-similarity relevance, effectively suppressing noisy candidates. Extensive experiments on Twitter-15 and Twitter-17 demonstrate that MTGLWS consistently surpasses state-of-the-art text-only and multimodal baselines in both accuracy and macro-F1, while ablation studies verify the complementary value of caption augmentation and label-word filtering, especially under few-shot conditions. These findings highlight the efficacy of prompttemplate generation and controlled label-word selection in activating knowledge within large pre-trained models for multimodal sentiment understanding.