機械翻訳クイズ

クイズに答えると、答えと解説が表示されます。

1. What does parallel training data contain?
   • Source passages paired with corresponding translations
   • Unrelated documents grouped by font
   • Only target-language references
   • A list of model predictions without sources
2. What can mismatched source-target alignment teach a translation model?
   • A guaranteed improvement in fluency
   • The wrong relationship between passages
   • A correct translation for every language
   • A better BLEU reference
3. Why keep related documents together when splitting evaluation data?
   • To remove the need for references
   • To make tokenization identical
   • To reduce near-duplicate leakage across splits
   • To increase sentence length
4. A tokenizer splits one writing system into many more units. What should evaluators check?
   • Whether source and target use the same alphabet
   • Only the model’s name
   • Whether BLEU is always high
   • Length limits and whether either passage is truncated
5. What does BLEU compare?
   • Whether a translation is legally valid
   • Patterns of word sequences with reference translations
   • The full meaning and readability of a passage
   • The number of languages in a model
6. A candidate translation changes “did not arrive” to “did arrive.” What error should reviewers flag?
   • An insertion that cannot matter
   • A casing normalization choice
   • Incorrect negation that reverses meaning
   • A target-language tokenization win
7. A model has a strong average score across several well-resourced languages. What could that conceal?
   • Poor performance in a less-represented language or specialized domain
   • That all language pairs are equally reliable
   • That there are no terminology errors
   • That the test set has no duplicates
8. What should accompany an automatic metric in translation evaluation?
   • A single score without settings
   • Only a larger reference corpus
   • An assumption that meaning is preserved
   • A documented setup and bilingual review of specific errors