A Hybrid Machine Learning Framework for Sentiment-Based Early Crisis Detection in News Streams
International Journal of Global Information Systems · 04.10.2026 · 2 (1) · 134-161
DOI: 10.66395/globeis.21 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Global Information Systems, online erschienen am 04.10.2026. Verfasst von Sena Nur Çulcu und 1 weiteren Autor:innen. Die Arbeit nennt 24 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
The rapid growth of digital news sources has made the automatic analysis of news content increasingly important for the early identification of potential societal crises. In this study, an integrated approach is proposed to classify the sentiment of news headlines and to support the identification of potential crisis periods by monitoring temporal changes in the intensity of negative sentiment. News headlines from the News Category Dataset V3 were categorized into three sentiment classes—positive, neutral, and negative—using the VADER sentiment analysis method. For text representation, TF-IDF features, which capture the statistical importance of words, and MiniLM embeddings, which model semantic and contextual relationships, were employed. A hybrid representation was constructed by combining these two feature groups, and several machine learning models were comparatively evaluated on this representation. The best-performing model, a hybrid Linear SVM combining TF-IDF and MiniLM features, achieved 91.22% accuracy and a Macro-F1 score of 0.9099 in predicting the VADER-generated sentiment labels. Furthermore, when comparing VADER labels with independent human evaluations, the model achieved 86.67% accuracy, a Macro-F1 score of 0.8659, and a Cohen’s Kappa value of 0.7975. The resulting hybrid representation was then integrated into the proposed temporal crisis assessment framework, enabling sentiment classification outputs to monitor unusual changes in negative sentiment over time. The proposed approach not only classifies news headlines by sentiment but also enables analysis of temporal changes in the proportion of negative sentiment. By integrating the statistical strength of TF-IDF with the contextual representation capability of MiniLM, the proposed framework provides a high-performing, lightweight, and practical solution. It is anticipated that this study will contribute to decision support systems for monitoring news streams and facilitating the early assessment of potential crisis trends.
Zitieren
APA
Çulcu, S. N., & Koç, C. (2026). A Hybrid Machine Learning Framework for Sentiment-Based Early Crisis Detection in News Streams. International Journal of Global Information Systems, 2(1), 134-161. https://doi.org/10.66395/globeis.21
Vancouver
Çulcu SN, Koç C. A Hybrid Machine Learning Framework for Sentiment-Based Early Crisis Detection in News Streams. International Journal of Global Information Systems. 2026;2(1):134-161. doi:10.66395/globeis.21
BibTeX
@article{ulcu2026,
author = {Çulcu, Sena Nur and Koç, Canan},
title = {A Hybrid Machine Learning Framework for Sentiment-Based Early Crisis Detection in News Streams},
journal = {International Journal of Global Information Systems},
year = {2026},
volume = {2},
number = {1},
pages = {134--161},
doi = {10.66395/globeis.21}
}
EUVIMED