What question did this study set out to answer?

To analyze the impact of corpus sequencing on low-data neural machine translation (NMT) performance.

April 19, 2026Open Access

Ordering Dominance in Low-Data NMT Regimes: Corpus Sequencing Outperforms Contradiction Magnitude (Part I: Structural Ablation Evidence)

Key Points

To analyze the impact of corpus sequencing on low-data neural machine translation (NMT) performance.
Utilized a low-data EN→UA NMT protocol with a controlled dataset
Conducted a single-factor ablation study by altering only corpus sequencing
Held model, tokenizer, LoRA, seeds, and corpus content constant for comparison
Corpus sequencing was identified as a dominant factor affecting chain-level coherence
Reproducible degradation in coherence was noted, with boundary amplification around 11.6%
Contradiction magnitude did not show significant influence under the same conditions

Abstract

This work demonstrates that, under a fully controlled low-data EN→UA NMT protocol (N = 142), deterministic corpus ordering acts as the dominant structural driver within the tested regime. In a single-factor ablation with model, tokenizer, LoRA, seeds, and corpus content held constant, altering only corpus sequencing produces reproducible degradation in chain-level coherence, including boundary amplification (~11.6%), while entropy remains invariant. Under the same protocol, contradiction magnitude (ΔV) does not emerge as a dominant driver. Corpus organization is an active structural variable in low-data NMT — not a preprocessing detail. This finding motivates the layered diagnostic architecture described in the companion series. All results are strictly bounded to the fixed low-data protocol. No generalization beyond this setting is claimed.

Read Full Paperexternally

Demander à l'IA

Bookmark

View Full Paper