From e7d0be228874c484bc7ca44bd3e7f03f447820ec Mon Sep 17 00:00:00 2001 From: Clemens Neudecker Date: Wed, 20 Nov 2019 00:44:04 +0100 Subject: [PATCH] Update Preprocessing.md --- docs/Preprocessing.md | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/docs/Preprocessing.md b/docs/Preprocessing.md index b9b04e3..bb91fe6 100644 --- a/docs/Preprocessing.md +++ b/docs/Preprocessing.md @@ -11,11 +11,11 @@ Layout Analysis & Textline Extraction @[sbb_pixelwise_segmentation](https://gith ### OCR & Word Segmentation OCR is based on [OCR-D](https://github.com/OCR-D)'s [ocrd_tesserocr](https://github.com/OCR-D/ocrd_tesserocr) which requires [Tesseract](https://github.com/tesseract-ocr/tesseract) **>= 4.1.0**. The [Fraktur_5000000](https://ub-backup.bib.uni-mannheim.de/~stweil/ocrd-train/data/Fraktur_5000000/) model, which is trained on [GT4HistOCR](https://github.com/tesseract-ocr/tesstrain/wiki/GT4HistOCR) is used. -corpiu -The [PAGE-XML](https://github.com/PRImA-Research-Lab/PAGE-XML) produced by the [Layout Analysis & Textline Extraction](https://github.com/qurator-spk/neath/blob/master/docs/Preprocessing.md#layout-analysis--textline-extraction) is taken as input, and the output is [PAGE-XML](https://github.com/PRImA-Research-Lab/PAGE-XML) containing the text recognition results with absolute pixel coordinates describing bounding boxes for words. ### Tokenization ### Named Entity Recognition -For Named Entity Recognition, a [BERT-Base](https://github.com/google-research/bert) model was trained. [sbb_ner](https://github.com/qurator-spk/sbb_ner) is using a combination of unsupervised training on a large (~2.3m pages) OCR corpus in combination with supervised training on a small (50k tokens) annotated corpus. Further details are available in the [paper](https://corpora.linguistik.uni-erlangen.de/data/konvens/proceedings/papers/KONVENS2019_paper_4.pdf). +For Named Entity Recognition, a [BERT-Base](https://github.com/google-research/bert) model was trained for noisy OCR texts with historical spelling variation. + +[sbb_ner](https://github.com/qurator-spk/sbb_ner) is using a combination of unsupervised training on a large (~2.3m pages) [corpus of German OCR](https://zenodo.org/record/3257041) from the Digital Collections of the Berlin State Library in combination with supervised training on a small (47k tokens) [annotated corpus](https://github.com/EuropeanaNewspapers/ner-corpora/tree/master/enp_DE.sbb.bio) of OCR from digitized historical newspapers of the Berlin State Library. Further details are available in the [paper](https://corpora.linguistik.uni-erlangen.de/data/konvens/proceedings/papers/KONVENS2019_paper_4.pdf).