Klasifikasi Berita Palsu Politik Berbasis Distilbert Dengan Peningkatan Data Melalui Teknik Augmentasi Teks

Abstract

The proliferation of political fake news stands as a primary challenge in the digital era. Such news is intentionally crafted to mislead and influence public opinion, spreading rapidly through social media, thereby significantly impacting public discourse and democratic integrity. This research aims to evaluate the effectiveness and efficiency of DistilBERT, comparing it with BERT and RoBERTa, and to analyze the impact of text augmentation as a data class balancing strategy. The methodology involves secondary data collection from the LIAR-PLUS dataset, text preprocessing, and the application of augmentation techniques such as synonym replacement, random insertion, and back translation. The DistilBERT model was trained in two scenarios: without and with data augmentation, with hyperparameter optimization performed using Optuna. Model performance was evaluated using standard classification metrics, including inference time. Results indicate that RoBERTa, after optimization, is the most effective model, achieving an f1-Score of 0,7761 and an inference time of 3,98 seconds, excelling in detecting the minority class. DistilBERT emerged as the most efficient model with an f1-Score of 0,6971 and an inference time of 2,29 seconds, while BERT also demonstrated improvement, reaching an f1-Score of 0.7349 with an inference time of 3,83 seconds. Among the augmentation techniques, back translation was the most effective in improving fake news detection (f1-score 0,6810), though it performed slightly below DistilBERT without augmentation. This research contributes to the development of more effective political fake news classification systems.