Neural OCR Post-Hoc Correction of Historical Corpora

Abstract Optical character recognition (OCR) is crucial for a deeper access to historical collections. OCR needs account orthographic variations, typefaces, or language evolution (i.e., new letters, word spellings), as the main source of character, word, segmentation transcription errors. For digital corpora prints, errors are further exacerbated due low scan quality and lack standardization. task post-hoc correction, we propose neural approach based on combination recurrent (RNN) deep convolutional network (ConvNet) correct At level flexibly capture errors, decode corrected output novel attention mechanism. Accounting input similarity, loss function that rewards model’s correcting behavior. Evaluation book corpus in German shows our models robust capturing diverse reduce error rate 32.3% by more than 89%.