Low-resource Post Processing of Noisy OCR Output for Historical Corpus Digitisation

Low-resource Post Processing of Noisy OCR Output for Historical Corpus Digitisation Caitlin Richter, Matthew Wickes, Deniz Beser, Mitch Marcus University of Pennsylvania Departments of Linguistics and Computer and Information Science, 3330 Walnut Street, Philadelphia PA 19104 {ricca, wickesm, dbeser}@seas.upenn.edu, mitch@cis.upenn.edu Abstract We develop a post-processing system to efficiently correct errors from noisy optical character recognition (OCR) in a 2.7 million word Faroese corpus. 7.6% of the words in the original OCR text contain an error; fully manual correction would take thousands of hours due to the size of the corpus. Instead, our post-processing method applied to the Faroese corpus is projected to reduce the word error rate to 1.3% with around 65 hours of human annotator work. The foundation for generating corrected text is an HMM that learns patterns of OCR error and decodes noisy OCR character sequences into hypothesised correct language. A dictionary augments the HMM by contributing additional language knowledge, and a human annotator provides judgements in a small subset of cases that are identified as otherwise most prone to inaccurate output. An interactive workstation facilitates quick and accurate input for annotation. The entire toolkit is written in Python and is being made available for use in other low-resource languages where standard OCR technology falls short of desirable text quality. Supplementary analyses explore the impact of variable language resource availability and annotator time limitations on the end quality achievable with our toolkit. Keywords: Faroese, OCR, low resource corpus development, human-assisted post processing 1. Introduction 1.1. Motivation We provide a toolkit to improve unsatisfactory text qual- ity in a newly digitised Faroese corpus (Føroyamálsdeildin, nd). The 2.7 million word corpus contains historical inter- views of unique linguistic and cultural value, which would provide valuable material for research in a variety of dis- ciplines if it were readily accessible (e.g. Heycock and Wallenberg, 2013; Galbraith, 2016). The interviews were recorded on tape up to 50 years ago, transcribed by type- writer, and later scanned (Figure 1) before being input to unknown commercial optical character recognition (OCR) software for text extraction. Figure 1: Partial scanned typewriter page of Faroese inter- view transcription. This process significantly corrupted the text with incorrect recognition. A lack of enough language-specific data for good Faroese language models in the initial OCR process had a large role in damaging the text quality; many of the errors produce tokens with distinctly un-Faroese letter se- quences, or even numbers and punctuation in the middle of words (like OCR error form r!ey^ur for deyður). Close to 8% of tokens (words) contain at least one error, suggesting a rough expectation for on average every sentence to con- tain one incorrectly recognised (non-)word. This error rate severely degrades the usability of the corpus for NLP, com- putational linguistic research, or any interest that relies on basic text search to identify relevant documents within the 2.7 million word corpus. Faroese is spoken by only about 50,000 people on the Faroe Islands, located between Norway, Scotland, and Ice- land (Thráinsson et al., 2012). It …

Content truncated to 500 words