Generating a Training Corpus for OCR Post-Correction Using Encoder-Decoder Model

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 1006–1014, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Generating a Training Corpus for OCR Post-Correction Using Encoder-Decoder Model Eva D’hondt LIMSI, CNRS, Univ. Paris-Saclay F-91405 Orsay eva.dhondt@gmail.com Cyril Grouin LIMSI, CNRS, Univ. Paris-Saclay F-91405 Orsay cyril.grouin@limsi.fr Brigitte Grau LIMSI, CNRS, ENSIIE, Univ. Paris-Saclay F-91405 Orsay bg@limsi.fr Abstract In this paper we present a novel ap- proach to the automatic correction of OCR-induced orthographic errors in a given text. While current systems depend heavily on large training corpora or exter- nal information, such as domain-specific lexicons or confidence scores from the OCR process, our system only requires a small amount of relatively clean train- ing data from a representative corpus to learn a character-based statistical language model using Bidirectional Long Short- Term Memory Networks (biLSTMs). We demonstrate the versatility and adaptabil- ity of our system on different text corpora with varying degrees of textual noise, in- cluding a real-life OCR corpus in the med- ical domain. 1 Introduction Recently, Optical Character Recognition (OCR) technology has improved substantially, which has allowed for a large-scale digitization of textual resources such as books, old newspapers, an- cient hand-written documents, etc. (Romero et al., 2011). The quality of the subsequent digital cor- pora can vary substantially, depending on factors such as quality of the original paper, ink quality, differences in fonts, etc. The amount of noise in digital collections can have a severe negative im- pact on the accuracy of subsequent text mining processes (e.g., Named Entity Recognition, Infor- mation Extraction, etc.). OCR post-correction techniques are used to improve the text quality of OCR output. Tra- ditionally, they rely on domain-specific lexi- cons (de Does and Depuydt, 2013) and character- based errors statistics obtained from a corrected training set (Kumar and Lehal, 2016). However, they have some drawbacks that limit their useful- ness for specific, low-resource domains. Such re- sources are expensive to create and for highly spe- cialized texts (e.g., medical domain) not always possible to obtain. The recent advances in neu- ral network models, based on textual context and needing no external resources, provide new oppor- tunities for OCR post-correction. Character-level sequence modeling architectures are especially suited for this task (Chrupała, 2014; Schmaltz et al., 2016), as they reduce the complexity at output time. Moreover, current systems are of- ten limited to processing texts with a limited de- gree of OCR corruption, i.e., so-called single-error word corrections (Kissos and Dershowitz, 2016) and correction of OCRed corpora that have been generated by older OCR engines can prove too challenging. The correct recognition of historical texts remains an open challenge (Kluzner et al., 2009). A general-purpose OCR post-correction tool should be adaptable to the ratio of error that is present in the OCR output in order to deal with both types of errors. In this paper, we propose a novel approach to OCR post-correction using bidirectional re- current neural networks for learning a robust character-based language model that (i) …

Content truncated to 500 words