Comparative Evaluation of IndoBERT-Based Architectures for Imbalanced Indonesian News Title Classification

Abstract

Stacking multiple imbalance-mitigation techniques on top of a pretrained transformer is widely assumed to compound their individual benefits, yet rigorous component-wise evidence for this assumption remains scarce in the Indonesian text classification literature. Four classification architectures are compared in this work on a publicly available Indonesian news title corpus. The working set contains 27,266 short headlines, drawn as a 30% stratified subsample from a cleaned corpus of 90,891 headlines, spread over nine target categories with a class ratio of 13.29. Three reference architectures are constructed: an LSTM trained from scratch with Random Oversampling, a bidirectional LSTM augmented with additive attention, and a fine-tuned IndoBERT on the oversampled training partition. A fourth architecture extends IndoBERT through three additions, namely learned attention pooling over contextual token embeddings, focal modulation applied on top of the cross-entropy term, and minority-class paraphrasing via Indonesian–English–Indonesian back-translation. Every configuration is evaluated through stratified 5-fold cross-validation, paired t-tests with Bonferroni correction across three comparisons, and McNemar tests on the held-out partition. The fine-tuned IndoBERT with Random Oversampling alone reaches the highest macro F1 of 0.837. By contrast, the combined configuration drops to 0.799, and statistical verification confirms that the gap is systematic rather than attributable to fold-level variation. A component-wise ablation isolates focal modulation as the principal driver of the decline, because it disturbs an already-balanced training distribution. The principal outcome of this study is empirical evidence indicating that composing several imbalance-oriented techniques on a pretrained transformer can yield adverse interactions rather than cumulative gains.