Leveraging Text Repetitions and Denoising Autoencoders in OCR Post-correction

TL;DR

This paper introduces a novel OCR post-correction method that leverages text repetitions and synthetic training data to improve accuracy without requiring manually corrected training pairs, demonstrating significant improvements on historical Finnish newspapers.

Contribution

The study presents a new approach to OCR post-correction that estimates errors from repeated text spans and trains a neural model with synthetic data, reducing reliance on manual annotations.

Findings

Significant OCR accuracy improvement over baseline systems.

Synthetic training data effectively captures error distribution.

Method outperforms models using uniform noise generation.

Abstract

A common approach for improving OCR quality is a post-processing step based on models correcting misdetected characters and tokens. These models are typically trained on aligned pairs of OCR read text and their manually corrected counterparts. In this paper we show that the requirement of manually corrected training data can be alleviated by estimating the OCR errors from repeating text spans found in large OCR read text corpora and generating synthetic training examples following this error distribution. We use the generated data for training a character-level neural seq2seq model and evaluate the performance of the suggested model on a manually corrected corpus of Finnish newspapers mostly from the 19th century. The results show that a clear improvement over the underlying OCR system as well as previously suggested models utilizing uniformly generated noise can be achieved.