Short-Answer Exact Extraction
Use for exact-match short-answer tasks with a minimal slot output.
Confirm exact-match scoring; a bare/declarative prompt asks for the slot value, not Yes/No.
Identify the slot from the prompt form:
- Bare entity-only prompt: the entity is given; return the informative predicate value from its definitional sentence (country/location/category), not an appositive official/expanded name.
- Truncated fragment: return the matching literal continuation span, not a source/title.
- Quoted slogan/catchphrase/title: return only the minimal associated title/entity, never a sentence that merely contains it.
- "this play/work" + setting clue: return the work's title, not the location.
- Definition/category: return the exact defining-sentence noun phrase with article/inflection.
Anchor to the strongest explicit context support, not prominence/prior knowledge:
- Multiple candidates: choose the entity explicitly linked to the queried facts; ignore "current" unless context links it.
- Quoted fragments: leading ellipsis -> preceding token; truncated end -> immediate continuation.
Return only the minimal slot phrase with the expected surface form:
- Publication names: return the full title as given, including the type noun ("Forbes magazine", not "Forbes").
- Diacritics: use the unaccented ASCII form when context supports it; never restore diacritics absent from the context.
- Nutrients: bare letter ("C", not "Vitamin C").
- Term variants: shortest base term ("a balloon", not "speech balloon").
- Named places: keep leading article, omit added category noun ("the Sahara", not "Sahara Desert"); keep official prefix ("Lake X").
- No parentheticals/expanded names; never emit a sentence containing the answer.
Verify against prompt and strongest context span; match grammar/capitalization and singular/plural; return no punctuation/explanation/labels.