Text-Only Caption Sentiment: A Balanced, Calibrated Comparison of RoBERTa and TextCNN
Journal of Computational Science and Applications · 11.12.2025 · 2 (2) · 38-46
DOI: 10.51846/jcsa.v2i2.4589 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Computational Science and Applications, online erschienen am 11.12.2025. Verfasst von Saba Ramzan und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Short-text sentiment analysis for social-media captions is challenging due to lexical sparsity, domain drift, and neutral-class dominance. We present a compact text-only pipeline that contrasts a contextual transformer (RoBERTa-base) with a lightweight convolutional baseline (TextCNN) under a shared three-way label space positive, neutral, negative. To ensure fair, reproducible comparisons in imbalanced settings, we pair stable fine-tuning with class-balanced loss, report macro-averaged Precision/Recall/F1 alongside per-class scores and confusion matrices, and quantify probability reliability via temperature scaling. We further outline deployment pathways for constrained environments using parameter-efficient fine-tuning (e.g., LoRA) and post-training quantization. On held-out caption data, RoBERTa consistently surpasses TextCNN most notably on the neutral class while the proposed training recipe reduces seed sensitivity and improves threshold reliability after calibration. The results position RoBERTa as an accuracy-oriented default for caption sentiment and validate TextCNN as a transparent, CPU-friendly baseline; under edge constraints, PEFT and quantization retain most transformer gains at substantially lower cost. The study provides a practical, reproducible template models, metrics, and deployment guidance for researchers and practitioners seeking dependable caption-level sentiment systems without multimodal components.
Zitieren
APA
Ramzan, S., & Ali, J. (2025). Text-Only Caption Sentiment: A Balanced, Calibrated Comparison of RoBERTa and TextCNN. Journal of Computational Science and Applications, 2(2), 38-46. https://doi.org/10.51846/jcsa.v2i2.4589
Vancouver
Ramzan S, Ali J. Text-Only Caption Sentiment: A Balanced, Calibrated Comparison of RoBERTa and TextCNN. Journal of Computational Science and Applications. 2025;2(2):38-46. doi:10.51846/jcsa.v2i2.4589
BibTeX
@article{Ramzan2025,
author = {Ramzan, Saba and Ali, Junaid},
title = {Text-Only Caption Sentiment: A Balanced, Calibrated Comparison of RoBERTa and TextCNN},
journal = {Journal of Computational Science and Applications},
year = {2025},
volume = {2},
number = {2},
pages = {38--46},
doi = {10.51846/jcsa.v2i2.4589}
}
EUVIMED