Span Annotation Adjudication Workbench

How to use Span Annotation Adjudication Workbench

Resolve annotation disagreements without guessing which annotator is right. Exact agreements are retained automatically; overlapping or one-sided ranges form review groups where you explicitly choose set A, set B or omission. Every selected annotation keeps its source ID, quote and record number.

1. Paste the exact source text and two flat annotation arrays with scalar offsets and matching quotes.
2. Run JSON review, inspect each conflict group and enter your own A, B or omit decisions for its current group ID.
3. Run again and export the resolved set only after ready is true; keep the source hash and decision evidence together.

Good to know

• One well-formed Unicode text of at most 20,000 UTF-16 characters; offsets count Unicode scalar values in logical source order, with zero-based inclusive start and exclusive end. No normalization or grapheme-boundary adjustment is performed. Every quote must match the current text exactly.
• Each set has at most 100 unique-ID, nonempty, nonoverlapping spans. Nested or overlapping spans within one set are refused. Original JSON numeric spellings must represent exact integers; rounded fractions are refused. Touching ranges belong to separate groups.
• Only identical range/label pairs agree automatically. Each other overlap-connected group needs a current explicit side or omit decision. JSON can contain unresolved partial selections; CSV requires all conflicts resolved. No accuracy statistic, ontology reconciliation or automatic boundary repair is claimed.
• The downloaded report includes source text, both annotation sets, SHA256 of exact UTF-8 text and human decisions. It can contain private information and is a custom audit profile, not a complete Web Annotation interchange document.
• Processing is local. Output is limited to 1,000,000 text characters unless a stricter file limit is stated.