Session 11
Text recognition: introduction to eScriptorium
Hodel Taught by Tobias Hodel.
In this session
eScriptorium, hands-on: from a page image to machine-readable text.
OCR and HTR
- Optical Character Recognition and Handwritten Text Recognition: the same pipeline, different difficulty
- Why printed early modern text is often harder than neat nineteenth-century handwriting
- The pipeline: image → binarisation → segmentation → transcription → export
- Kraken as the engine underneath eScriptorium
Segmentation
- Regions and lines, and the baseline model
- Reading order, which is a scholarly decision as much as a technical one
- Correcting segmentation by hand — and why an hour here saves a day later
- Layout that breaks models: marginalia, tables, multi-column pages, insertions
Transcription and ground truth
- Transcribing lines in the editor
- Transcription guidelines: abbreviations,
u/v,i/j, long s, capitalisation. Decide before you start, write it down, and stay consistent — this is an editorial decision, exactly as in the TEI session - How much ground truth a model needs
- Training and fine-tuning an existing model rather than starting from scratch
Evaluation
- CER (character error rate) and WER (word error rate): what they measure
- Why a 5% CER can be excellent or useless depending on what you want to do with the text
- Where the errors concentrate — and they always concentrate somewhere
- Ground truth is not truth: it is your transcription decisions, made consistent
NoteThis is the Intro’s machine-learning session, from the inside
The Intro asks where bias enters a machine learning pipeline. Here you occupy the position that answer names: you are the person producing the labels.
Export
- ALTO and PAGE XML, and what they contain besides the text
- Getting from ALTO to TEI
- Publishing images and text together with IIIF
TipIn parallel in the Intro