BLEU (Bilingual Evaluation Understudy), proposed by Kishore Papineni et al. in 2002, computes a comprehensive score by calculating the n-gram precision (typically 1-gram to 4-gram) between the candidate and reference texts and multiplying it by a brevity penalty factor. The score ranges from 0 to 100, where higher values indicate better performance.

ROUGE (Recall-Oriented Understudy for Gisting Evaluation), proposed by Chin-Yew Lin in 2004, was originally used for automatic summarization evaluation and has since been extended to machine translation and text generation tasks. ROUGE-L calculates the harmonic mean of recall and precision based on the Longest Common Subsequence (LCS), considering both content coverage and word order. It evaluates the fidelity of generated text based on n-gram overlap with reference texts, making it suitable for open-ended generation with reference answers. Higher values indicate better performance. This metric is sensitive to the coverage of key information in summaries and is one of the standard choices for text summarization evaluation.

Both metrics are used for open-ended generation with reference answers.