Abstract
Optical Character Recognition (OCR) enables the digitisation and analysis of large document collections, but recognition errors remain a challenge, particularly for historical documents. This paper presents TOOL (Treating OCR Output as a Language), a modular approach that formulates OCR post-correction as a sequence-to-sequence translation task. A fine-tuned MarianMT encoder–decoder model transforms noisy OCR output into corrected text without modifying the underlying recognition engine. We evaluate TOOL on a collection of historical German vocational education and training documents spanning the period from 1871 to the present day. On 139 held-out test documents, TOOL reduces mean word error rate from 0.792 for Tesseract alone to 0.600, corresponding to a relative reduction of approximately 24.3%. Improvements are also observed in Jaccard bigram similarity, BLEU-4, METEOR, and ROUGE-2, whereas mean normalised Levenshtein distance increases. Comparisons with Hunspell, Phi-4, and Qwen2.5-VL show that the larger language and vision-language models achieve higher accuracy on several metrics. However, TOOL offers a competitive quality–runtime trade-off, with an average end-to-end processing time of 30.12 s per document, compared with 96.57 s for Qwen Transcribed. The results support the feasibility of translation-based OCR post-correction as an additional processing stage in existing OCR workflows, while generalisation to other document collections, OCR engines, and languages remains to be investigated.
Funding Information
  • Deutsche Forschungsgemeinschaft (90893796)

This publication has 22 references indexed in Scilit: