chrF
chrF is an automatic translation-quality metric that compares output to a reference at the character level rather than the whole-word level. That makes it fairer to languages where a single word carries what English spreads across several, and more forgiving of inflection.
How it works
chrF compares the machine output and the reference by the character sequences they share, not whole words. Because it works below the word, it gives partial credit when a translation gets the stem right but the ending different, which is common in languages with rich inflection.
Like other reference-based metrics it is automatic, cheap and repeatable, and like them it still rewards resemblance to one reference rather than understanding meaning.
How SourceTarget uses it
SourceTarget includes chrF in its composite quality score, weighting it more for morphologically rich languages where whole-word metrics like BLEU are unfairly harsh. It is one of several signals behind the final score, never the sole judge.
chrF compared with BLEU score
| chrF | BLEU score | |
|---|---|---|
| Compares at | Character level | Whole-word level |
| Fairer to | Rich word forms and inflection | Languages close to the reference wording |
| Shared limit | Still reference-based, not meaning-aware | Still reference-based, not meaning-aware |