Rare but Severe Neural Machine Translation Errors Induced by Minimal Deletion: An Empirical Study on Chinese and English

TL;DR

The paper investigates rare but severe translation errors induced by minimal deletions in in-domain English–Chinese neural machine translation using character-based models. It proposes an enumeration-based method to identify severe errors by deleting single tokens and classifying the resulting errors into four types. The study compares two data regimes (1M and 10M sentences) and both translation directions, revealing that word deletions are more likely to trigger severe errors than character deletions, and that more data reduces overall error probabilities while sometimes increasing sensitivity to perturbations. The findings have implications for robustness, error detection, and training strategies in Chinese–English MT, and motivate broader cross-linguistic analyses and detector development.

Abstract

We examine the inducement of rare but severe errors in English-Chinese and Chinese-English in-domain neural machine translation by minimal deletion of the source text with character-based models. By deleting a single character, we can induce severe translation errors. We categorize these errors and compare the results of deleting single characters and single words. We also examine the effect of training data size on the number and types of pathological cases induced by these minimal perturbations, finding significant variation. We find that deleting a word hurts overall translation score more than deleting a character, but certain errors are more likely to occur when deleting characters, with language direction also influencing the effect.

Figures (1)

• Figure 1: Zh-En bleu as function of characters removed on valid sentences with 95% confidence intervals. There is a linear relationship, with average bleu converging as more tokens are removed.