Overview of the 2017 ALTA Shared Task: Correcting OCR Errors

Overview of the 2017 ALTA Shared Task: Correcting OCR Errors Diego Moll´a and Steve Cassidy Department of Computing Macquarie University Sydney, Australia diego.molla-aliod@mq.edu.au steve.cassidy@mq.edu.au Abstract This paper presents an overview of the 8th ALTA shared task that ran in 2017. The task was to correct OCR errors from scans of newspapers stored in the Trove database maintained by the National Li- brary of Australia. We introduce the task, describe the data and present the results of the participating teams. 1 Introduction Many digital documents are the result of scanning printed copies. These documents, although in dig- ital form, are in fact images, and as such, stan- dard natural language processing techniques such as text search cannot be applied to them. The National Library of Australia1 maintains an archive of scanned Australian publications in the Trove database2. Many of these scans have been processed through Optical Character Recognition (OCR) and form a searchable resource with over 500 million items. But the OCR output may con- tain errors which need to be corrected. Trove has corrected the errors through a process of collabo- rative editing of the output of the OCR system. The goal of the 2017 ALTA Shared Task is to automatically correct errors of OCR from a sub- set of scans from the Trove database. Over 7,000 documents were downloaded from the Trove database. For each document, the original output of the OCR system was used as the input text to the shared task, and the corrected versions were used as the target text. A total of 6,000 documents and their corrected versions were provided as the training set, and the rest was used to evaluate the system results. 1https://www.nla.gov.au/ 2http://trove.nla.gov.au/ This paper is structured as follows. Section 2 describes the shared task. Section 3 briefly in- troduces related research on OCR. Section 4 de- scribes the data set that was used. Section 5 de- tails the evaluation process. Section 6 presents and discusses the results. Finally, Section 7 concludes this paper. 2 The 2017 ALTA Shared Task The 2017 ALTA Shared Task is the 8th of the shared tasks organised by the Australasian Lan- guage Technology Association (ALTA). Like the previous ALTA shared tasks, it is targeted at uni- versity students with programming experience. The general objective of these shared tasks is to in- troduce university students to the sort of problems that are the subject of active research in a field of natural language processing. There are no limitations on the size of the teams or the means that they can use to solve the prob- lem, as long as the processing is fully automatic — there should be no human intervention. As in past ALTA shared tasks, there are two cat- egories: a student category and an open category. • All the members of teams from the student category must be university students. The teams cannot have members that are full-time employed or that have completed a PhD. • Any other teams fall into the open category. …

Content truncated to 500 words