Sentiment-aware-image-captioning-vcsc-code-v1.0.zip
This Zenodo archive contains the code accompanying the manuscript “Sentiment-aware image captioning with visually consistent sentiment calibration.” The repository provides a consolidated and reproducible implementation of the proposed multimodal approach for generating sentiment-aware image captions. The framework integrates Vision Transformer (ViT) for visual feature extraction, a GPT-2–based decoder for caption generation, and multiple sentiment classification models including LSTM, GRU, CNN, Transformer, and fine-tuned BERT. In addition, the archive implements the proposed Visually Consistent Sentiment Calibration (VCSC) mechanism, which
Resource information
- Format
- zip
- Original repository
- 19367782
- DOI
- 10.5281/zenodo.19367782
- Credit basis
- Creators in Zenodo record metadata
- Access
- Open files available from the original Zenodo deposit
- Last checked
- Nov 5, 2026
- License
- mit-license
- Last record update
- Nov 5, 2026
About this resource
This Zenodo archive contains the code accompanying the manuscript “Sentiment-aware image captioning with visually consistent sentiment calibration.” The repository provides a consolidated and reproducible implementation of the proposed multimodal approach for generating sentiment-aware image captions. The framework integrates Vision Transformer (ViT) for visual feature extraction, a GPT-2–based decoder for caption generation, and multiple sentiment classification models including LSTM, GRU, CNN, Transformer, and fine-tuned BERT. In addition, the archive implements the proposed Visually Consistent Sentiment Calibration (VCSC) mechanism, which aligns textual sentiment with visual context through a lightweight post-hoc calibration strategy without requiring joint retraining. The package includes: - A unified Python implementation of the full pipeline (caption generation, sentiment classification, and calibration) - Baseline models for comparative evaluation (TF-IDF + classical ML, deep learning, and transformer-based methods) - Experimental configurations aligned with the manuscript - Documentation and usage instructions - Original notebooks for transparency and reproducibility The implementation supports five-class sentiment labeling with emoji-based representation and reproduces the experimental setup reported in the manuscript, including evaluation metrics (BLEU, ROUGE, METEOR, CIDEr for captioning; accuracy, precision, recall, and F1-score for sentiment classification) and ablation analysis of the VCSC component. This archive is intended to ensure full reproducibility of the results and to facilitate further research in multimodal sentiment-aware image captioning.