TREC-5 “Confusion” Track
3 versions of ‘94 Federal Register
- true text
- NIST OCR output (approx 5% error rate)
- NIST OCR output of downsampled image
(approx 20% error rate)
Evaluation: mean reciprocal rank of target
- equivalent to average precision since only 1 rel doc
See “Report on the TREC-5 Confusion Track” by Kantor and Voorhees in TREC-5 Proceedings (http://trec.nist.gov/)