For transformer models, oversampling and backtranslation augmentation improve performance while freezing the encoder substantially decreases it¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Both oversampling and data augmentation are associated with higher average transformer performance, with relative gains of 22% and 26% overall respectively, while freezing the backbone decreased overall performance by 12%. → Zsolt Csibi 2026

Evidence¶

Zsolt Csibi 2026¶

Zsolt Csibi, Kristian Fenech, Mónika Sándor, Mónika Serfőző, Kinga Gyöngy. (2026). Automatic Reflection Level Classification in Hungarian Student Essays. https://arxiv.org/abs/2605.02402 doi:10.48550/arxiv.2605.02402

q2 · i? · design · r2

Transformer ablation (Table 5) averaged over the 2 transformer models and hyperparameter settings, reporting ROS at 0.6029 overall (22%↑), backtranslation at 0.6412 (26%↑) and freeze encoder at 0.4786 (-12%↓) relative to baseline. The article states "freezing the backbone during training significantly decreased the performance".

"However, freezing the backbone during training significantly decreased the performance of the models, which shows the importance of full fine-tuning when adapting transformer models to specific tasks."

Discussion¶

Related Claims¶

• In the shallow-model ablation, class balancing lowered every metric, oversampling did not improve overall performance, and semantic embeddings raised ROC-AUC but lowered accuracy and F1 — related
• Shallow machine learning with SMOTE oversampling and Qwen3 embeddings achieves the best overall reflection-classification performance (0.7176) on Hungarian essays — related
• Hungarian transformer models achieve slightly lower overall score (about 68%) but better minority-class sensitivity than shallow models — related
• Removing unsupervised pre-training hurts performance across all tasks, causing a 14.8% decrease relative to the full model — related
• T-tests show no significant shallow-vs-transformer difference in accuracy, weighted F1 or macro-F1, but a significant ROC-AUC advantage for the shallow model — related