Reference translation
A reference translation is a human translation treated as the gold standard for evaluating machine output. Metrics like BLEU, chrF and reference-based COMET score a translation by how closely it matches one or more references. The quality of the evaluation depends on the reference: a single reference cannot represent every valid way to translate a sentence.
How it works
To benchmark an engine, you hold a test set of source sentences with trusted human translations. The engine translates the source, and a metric compares its output against the references. Using several references per sentence helps, because it captures more of the legitimate variation in how a sentence can be translated.
The limitation is built in: any correct translation that happens not to resemble the reference is scored down. This is why reference-based metrics are best for comparing systems, not for passing final judgement on a single sentence.
How SourceTarget uses it
SourceTarget uses reference translations when benchmarking engines on a language direction: a held set of source sentences with trusted human translations lets metrics compare engines fairly. In live production, where no reference exists, reference-free quality estimation takes over.