← AI Terminology
Exact Match
Exact match (EM) scores 1 only when a predicted answer string equals the reference (after simple normalisation), else 0 — a strict accuracy for QA and extraction.
Common in SQuAD-style and closed-form evals.
Common in SQuAD-style and closed-form evals.
Why It Matters in AI
EM is simple and unambiguous but brittle to phrasing. Still foundational for extractive QA, and often paired with F1 token overlap. Knowing when EM is too harsh guides metric choice for generative systems.
Key Points
| Aspect | Description |
|---|---|
| Use | SQuAD EM, many closed-form benchmarks |
| Rule | Normalised string equality → 1/0 |
| Norms | Lowercase, strip articles/punctuation (task-specific) |
| Limits | Synonyms and reorders fail EM |
| Related | Accuracy, pass@k for code |
| Pair with | Token F1, semantic similarity, LLM judges |
Simple Analogy
A spelling bee with no partial credit — the answer is either exactly right or wrong.
Common Usage Examples
- SQuAD EM/F1 reporting
- Extractive QA pipelines optimised for EM
- Don’t use EM alone for free-form chat
- Normalisation scripts must be consistent
Summary
In short: Exact match awards credit only for fully identical answers — a strict, simple metric for QA and extraction tasks.