A Generative Probabilistic OCR Model for NLP Applications

A Generative Probabilistic OCR Model for NLP Applications Okan Kolak Computer Science and UMIACS University of Maryland College Park, MD 20742, USA okan@umiacs.umd.edu William Byrne CLSP The Johns Hopkins University Baltimore, MD 21218, USA byrne@jhu.edu Philip Resnik Linguistics and UMIACS University of Maryland College Park, MD 20742, USA resnik@umiacs.umd.edu Abstract In this paper, we introduce a generative prob- abilistic optical character recognition (OCR) model that describes an end-to-end process in the noisy channel framework, progressing from generation of true text through its transforma- tion into the noisy output of an OCR system. The model is designed for use in error correc- tion, with a focus on post-processing the output of black-box OCR systems in order to make it more useful for NLP tasks. We present an implementation of the model based on finite- state models, demonstrate the model’s ability to significantly reduce character and word er- ror rate, and provide evaluation results involv- ing automatic extraction of translation lexicons from printed text. 1 Introduction Although a great deal of text is now available in elec- tronic form, vast quantities of information still exist pri- marily (or only) in print. Critical applications of NLP technology, such as rapid, rough document translation in the field (Holland and Schlesiger, 1998) or information retrieval from scanned documents (Croft et al., 1994), can depend heavily on the quality of optical character recog- nition (OCR) output. Doermann (1998) comments, ”Al- though the concept of a raw document image database is attractive, comprehensive solutions which do not require complete and accurate conversion to a machine-readable form continue to be elusive for practical systems.” Unfortunately, the output of commercial OCR systems is far from perfect, especially when the language in ques- tion is resource-poor (Kanungo et al., in revision). And efforts to acquire new language resources from hardcopy using OCR (Doermann et al., 2002) face something of a chicken-and-egg problem. The problem is compounded by the fact that most OCR system are black boxes that do not allow user tuning or re-training — Baird (1999, re- ported in (Frederking, 1999)) comments that the lack of ability to rapidly retarget OCR/NLP applications to new languages is “largely due to the monolithic structure of current OCR technology, where language-specific con- straints are deeply enmeshed with all the other code.” In this paper, we describe a complete probabilistic, generative model for OCR, motivated specifically by (a) the need to deal with monolithic OCR systems, (b) the fo- cus on OCR as a component in NLP applications, and (c) the ultimate goal of using OCR to help acquire resources for new languages from printed text. After presenting the model itself, we discuss the model’s implementation, training, and its use for post-OCR error correction. We then present two evaluations: one for standalone OCR correction, and one in which OCR is used to acquire a translation lexicon from printed text. We conclude with a discussion of related research and directions for future work. 2 The Model Generative “noisy channel” models relate an observable string …

Content truncated to 500 words