Manuscript Code: Sentiment-aware Image Captioning with Visually Consistent Sentiment Calibration (VCSC)
This repository provides the manuscript code for the research work: “Sentiment-aware Image Captioning with Visually Consistent Sentiment Calibration (VCSC)”. The implementation presents a unified multimodal pipeline for generating sentiment-aware image captions. It integrates Vision Transformer (ViT) for visual feature extraction and a GPT-2-based decoder for caption generation (Vision–GPT framework). The framework further includes a comprehensive five-class sentiment classification module using: - TextBlob-based weak supervision - Classical machine learning baselines (TF-IDF with SVM, Random Forest, Decision Tree, Naive Bayes) - Deep learnin
Resource information
- Format
- py
- Original repository
- 19558277
- DOI
- 10.5281/zenodo.19558277
- 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 repository provides the manuscript code for the research work: “Sentiment-aware Image Captioning with Visually Consistent Sentiment Calibration (VCSC)”. The implementation presents a unified multimodal pipeline for generating sentiment-aware image captions. It integrates Vision Transformer (ViT) for visual feature extraction and a GPT-2-based decoder for caption generation (Vision–GPT framework). The framework further includes a comprehensive five-class sentiment classification module using: - TextBlob-based weak supervision - Classical machine learning baselines (TF-IDF with SVM, Random Forest, Decision Tree, Naive Bayes) - Deep learning models (LSTM, GRU, CNN, Transformer) - Fine-tuned BERT for contextual sentiment understanding In addition, the repository implements the proposed Visually Consistent Sentiment Calibration (VCSC) mechanism, a lightweight post-hoc calibration approach that aligns textual sentiment predictions with visual context without requiring retraining. The codebase is designed to ensure transparency and reproducibility. It includes: - Complete end-to-end implementation of the proposed pipeline - Experimental configurations aligned with the manuscript - Scripts for training, evaluation, and prediction - Exportable outputs including metrics, classification reports, and confusion matrices - Supporting documentation and reproducibility notes The implementation uses the publicly available Flickr30k dataset and supports flexible data loading configurations. This repository is intended to facilitate reproducibility, verification, and further research in sentiment-aware image captioning and multimodal learning.