The fastest way to lose trust in an AI system is to change it and not be able to say whether you helped or hurt. During my work at DataBytes I spent most of my time on exactly this problem: turning “the output seems fine” into a number you can defend.
Three metrics, three jobs
None of these is “the” answer — they measure different things, and you want more than one.
- ROUGE — overlap-based; good for summarisation, where covering the right content matters. Rewards recall of key phrases.
- BLEU — precision-oriented n-gram overlap; borrowed from translation, useful when phrasing should stay close to a reference.
- BERTScore — embedding-based; catches semantic similarity when the wording differs but the meaning matches. This is the one that saves you from over-penalising good paraphrases.
What actually mattered
- Run it at scale. I evaluated across 20,000+ samples. Small eval sets lie.
- Look at disagreement. Where the metrics disagree is where the interesting failures live — a high ROUGE, low BERTScore result usually means right words, wrong meaning.
- Automate it. Evaluation you have to run by hand is evaluation you’ll skip.
The point
Evaluation isn’t the boring bit you do at the end. It’s the thing that lets you improve a system on purpose instead of by accident — and it’s the difference between shipping AI and gambling with it.