Pengaruh Teknik Augmentasi Dalam Klasifikasi Berita Berbahasa Inggris Menggunakan Algoritma BERT

This study investigates the impact of data augmentation techniques on the performance of a BERT-based news classification model. Five experimental scenarios were evaluated: one without augmentation, three with individual augmentation techniques (Synonym Replacement, Back Translation, and Random Swap), and one combining all three methods. The model performance was assessed using classification metrics and cosine similarity analysis. Results show that Synonym Replacement provided the best improvement, achieving the highest F1-score of 0.984 and the lowest evaluation loss of 0.0489. However, several techniques generated augmented texts that were overly similar to the original data (cosine similarity > 0.99), limiting their effectiveness in increasing data diversity. This study concludes that the success of data augmentation largely depends on the semantic variety and quality of the newly generated data, which should enrich the training set without introducing noise.