Explain BLEU, ROUGE, and BERTScore. When would you use each?

All three compare generated text to a reference, but differently. BLEU measures n-gram precision — how much of the output appears in the reference — and was built for machine translation; it rewards exact word overlap. ROUGE measures n-gram recall — how much of the reference is captured — and is standard for summarisation. Both are cheap and reproducible but purely lexical: they punish valid paraphrases that use different words and can't judge meaning. BERTScore instead compares contextual embeddings of the two texts, so it credits semantic similarity even when the wording differs, at higher compute cost. Use BLEU/ROUGE for quick, comparable scoring where wording is constrained; BERTScore when paraphrase matters. For open-ended, subjective quality, none is sufficient — reach for LLM-as-a-judge or humans.