A RoBERTa-Driven Deep Feature Learning Framework for Real-Time Political Event Classification on Twitter

Abstract

The rapid expansion of social media platforms such as Twitter has enabled continuous real-time information exchange, with millions of posts generated daily that reflect ongoing real-world events, particularly political activities that dominate global discussions. However, relying on a purely manual approach for monitoring and classifying such high-volume textual data is inefficient, time-consuming, and inconsistent, making it unsuitable for real-time analysis and decision-making. To overcome these limitations, this study proposes an automated political event detection framework that integrates transformer-based representations with optimized classification strategies. The process begins with NLP preprocessing steps such as tokenization, stop-word removal, normalization, and lemmatization, followed by EDA to analyze data patterns and distributions. Context-aware semantic features are extracted using Lightweight Robustly Optimized Bidirectional Encoder Representations from Transformers (RoBERTa), enabling rich textual representations. In contrast to baseline models such as Stochastic Gradient Descent (SGD), Histogram-based Gradient Boosting (HGB), Greedy Tree Classifier (GTC), and Random Forest Classifier (RFC), the proposed approach incorporates a DNNbased feature selection mechanism combined with an optimized SGD classifier to enhance discriminative learning and effectively handle class imbalance. The system classifies posts into categories including disaster, political, positive, protest, riot, and terror. Experimental results demonstrate improved accuracy, scalability, and adaptability, highlighting the effectiveness of transitioning from manual analysis to an automated, intelligent system for real-time applications in governance, crisis response, and public safety monitoring.