Label Annotation through Biodiversity Enhanced Learning

Abstract

The LABELX (Label Annotation through Biodiversity Enhanced Learning) is an extension of the HERBIS NLP system reported previously (Heidorn & Wei, 2008). The objective of the system is to formaly structure output from Optical Character Recognition (OCR) of the highly variable labels of natural history museum specimens. OCR errors are common in the OCR output. Genus and species names are particularly prone to errors. Records are preprocessed using a fuzzy-match algorithm to find and replace genus and species names, including those with OCR errors, and replace those with a constant token. Integers and strings that begin with Alphabetic characters and end with numbers are also replaced with tokens. LABELX generates structured XML data and RDF and makes corrections to OCR errors in some fields. The main algorithm is a Hidden Markov Model (HMM). This poster reports an enhancement to the previous system with a larger data set.