Topic 7
Looking right versus being right
Every similarity metric rewards resemblance to a reference, not correctness. AI Engineering (Ch. 3, "Lexical similarity") collects the ways that goes wrong:
- Missing references. A correct output can score low simply because no reference looks like it. The book reports that Adept found its model Fuyu scoring poorly on some benchmark examples because correct answers were missing from the reference data.
- Wrong references. The organisers of the WMT 2023 Metrics shared task found many bad reference translations in their data. The book cites this as one reason reference-free metrics held their own against reference-based ones in matching human judgement (Freitag et al., 2023).
- Similarity isn't correctness. On the HumanEval code benchmark, OpenAI found that BLEU scores for correct and incorrect solutions were similar (Chen et al., 2021). A higher BLEU score doesn't mean code that works.
Predict first
Scroll sideways to see the whole diagram
If the output can be run and checked, such as code against tests, that's the score to use.
Text version
- If the output can be run and checked, such as code against tests, that's the score to use.
- If not, but there's one short right answer, compare exactly, after normalising case and spacing.
- If many phrasings are right and you have good references, compare by shared words or, better, by meaning.
- If none of those fit, use a judge with a written rubric.
- And check the metric itself: read a few high-scoring outputs and make sure they're actually good.
None of this makes reference metrics useless. It means each one fits some tasks, and every metric needs checking against outputs you've read yourself. When nothing exact fits, a judge with a clear rubric is the next step, and the next lesson.
Quick check