Machine translation automatically produces translated text. Its output can be compared with human reference translations to support faster iteration and system comparisons. section 2.0 section 2.1
BLEU measures lexical overlap, not meaning, synonyms, or paraphrases. Treat its score as a useful but limited signal of translation quality. section 2.0
Benchmarks and low-resource languages
BLEU remains a primary metric for comparing systems on WMT benchmarks. For languages such as Arabic and Telugu, token matching can provide a more reliable signal than an LLM judge. section 2.3.1 section 2.3.1











