What they are
- Require ground truth; described as "less useful but still worth a mention"
1:36 - Metrics like BLEU and ROUGE, originally designed for translation and summarization
1:36 - Work by comparing generated text to an ideal reference text (ideal summary/translation) and measuring word/symbol overlap, in a few more complicated ways
1:36
Limitation
- Useful in some scenarios, but overlap in symbols is often not very meaningful as a quality signal
1:58