Language Technologies Institute at Carnegie Mellon (LTI at CMU)
66 minutes
Quantitative measures of how good machine-generated text is: n-gram overlap metrics such as BLEU, embedding-based similarity such as cosine similarity, and the limitations of automatic metrics relative to human judgement.
Why this Concept is marked deprecated
Superseded by LLM-as-Judge Grading Calibrated Against Human Labels: BLEU-style n-gram overlap tracks human judgement poorly on open-ended generation, so professionals now use calibrated judges.