Unsupervised Multi-View Post-OCR Error Correction With Language Models

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8647–8652 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 8647 Unsupervised Multi-View Post-OCR Error Correction With Language Models Harsh Gupta† Luciano Del Corro† Samuel Broscheit‡∗ Johannes Hoffart† Eliot Brenner† †Goldman Sachs {Harsh.Gupta, Luciano.DelCorro, Johannes.Hoffart, Eliot.Brenner} @ gs.com ‡University of Mannheim broscheit@informatik.uni-mannheim.de Abstract We investigate post-OCR correction in a set- ting where we have access to different OCR views of the same document. The goal of this study is to understand if a pretrained language model (LM) can be used in an unsupervised way to reconcile the different OCR views such that their combination contains fewer errors than each individual view. This approach is motivated by scenarios in which unconstrained text generation for error correction is too risky. We evaluated different pretrained LMs on two datasets and found significant gains in realistic scenarios with up to 15% WER improvement over the best OCR view. We also show the im- portance of domain adaptation for post-OCR correction on out-of-domain documents. 1 Introduction Digital scans of printed paper are still one of the main sources of digitized text across industries, li- braries and governmental organizations. Scanned documents need to be processed by Optical Char- acter Recognition (OCR) systems in order to be consumed by natural language processing (NLP) pipelines. Unfortunately, OCR errors are perva- sive and input noise can severely hinder down- stream NLP applications, e.g., search (Stein et al., 2012), neural machine translation (Belinkov and Bisk, 2018) or NLU in general (Kumar et al., 2020). Prior work used text generation techniques or redundancy in similar passages for OCR error cor- rection, which is not appropriate in cases of low corpus redundancy or weak document contextual information. For example, this may pose a risk in sensitive documents in the legal or financial do- main, where documents tend to be templatized and specific information, such as legal entities or num- bers (e.g., interest rates or amounts), are specific to single documents and cannot be safely inferred. Other prior work uses multiple OCR views of a ∗ work done during an internship at Goldman Sachs OCR1: Total 0 ft suppl ies : 23 .64 OCR2: Total 0 % suppl tea .3 .64 Reconciled: Total 0 % suppl ies : 23 .64 Figure 1: Example ICDAR dataset. Output from two popular OCR systems (OCR1 and OCR2) and the rec- onciled version generated by our approach. document and reconciles them in a supervised way, requiring expensive and difficult to acquire training data. Therefore, we investigated post-OCR correction with multiple OCR views on the same document in an unsupervised way. A key assumption of this work is that different OCR views make mistakes in different parts of the input document. To create a better OCR view from multiple OCR inputs we use a language model (LM) to pick the most probable reconciliation. See Figure 1 for an example. The key question of this study is if it is possible to use a LM to reconcile …

Content truncated to 500 words