iConference2013: HERBIS/LABELX -- Machine Learning Approach to Parsing OCR Text

Tue, 2013-03-12 16:01 -- dpaul
TitleiConference2013: HERBIS/LABELX -- Machine Learning Approach to Parsing OCR Text
Publication TypePresentation
Year of Publication2013
AuthorsP. Heidorn, Bryan, and Zhang Qianjin
KeywordsAugmenting OCR, Darwin Core, Hidden Markov Models, Human Transcription, Machine Learning, N-Gramming, Naive Bayes, Natural Language, Parsing
AbstractA central goal of the digitization of museum specimen labels is to parse the content of the labels into standard fields such as those identified in the Darwin Core (DwC). Museum specimen labels number in the billions and may be hundreds of years old or created recently. They have some minimal layout consistency within subsets of labels but overall have a very high layout variability. The text generated via OCR from these labels tends to have a high rate of errors. This combination makes it extremely difficult for programmers to write rule systems or other matching algorithms to classify the sub-elements of the labels into the DwC fields. Statistical methods such as Naive Bayes, Hidden Markov Models and N-Gramming can be combined with human supervision and authority files to successfully parse OCR output into proper fields and correct some OCR errors based on context.