HTR Kennisuitwisseling 24 oktober 2023

De vierde bijeenkomst op 24 okbtober 2023

Bijeenkomst

Datum: 24 oktober 2023 Tijd: 11:00-13:00 Locatie: Spinhuis zaal 2.18, Oudezijds Achterburgwal 185, Amsterdam

Agenda

  • Rutger van Koert & Stefan Klut - Loghi
  • Carsten Schnober - HTR Quality

Aanwezig

  • Rutger van Koert
  • Niek Verhoeff
  • Rick Companje
  • Arno Bosse
  • Carsten Schnober
  • Gerhard de Kok
  • Milan van Lange
  • Kay Pepping
  • Stefan Klut
  • Marijn Koolen
  • Pauline van den Heuvel
  • Rik Hoekstra

Presentatie Loghi (Rutger van Koert en Stefan Klut)

  • Loghi:
  • LayPa:
  • Pre-processing
    • PageXML segmentation
    • semantic (type classificatie),
    • panoptic (groep classificatie o.b.v. pixels),
    • instance (o.b.v. polygonen)
  • Post-processing
    • semantic segmentation to baselines
    • cut out text lines (seam carving)
  • Next step:
    • transformer-based HTR
    • convolution+recurrent reads line-by-line, requires little training data
    • transformers require more training data
  • Other features
    • underline/strikethrough/superscript
    • finetuning of basemodels
    • PageXML 2019
    • data augmentation
  • Future
    • segmentation free OCR/HTR
    • generic models
    • LLM integration (including historical languages)
    • less GT needed hopefully

Presentatie HTR Quality (Carsten Schnober)

Classifying quality of digitzed documents

  • https://github.com/LAHTeR/htr-quality-classifier
  • developed for VOC documents in the GLOBALISE project
  • but applicable more generally
  • Context: LaHTeR project
    • faclitate manual and automatic analysis of HTR’ed documents
  • Challenges
    • downstream tasks have different requirements
    • Event detection is more sensitive than keyword searach to correct readign order and text segmentation
  • Ground truth
    • focus on layout analysis
    • good/medium/bad
    • 328 scans
  • Traffic light decision tree
    • red: re-do HTR
    • yellow: post-correction
    • green: done
  • scoring
    • Dictionary scores
    • trigram scores
    • garbage token score (heuristics e.g. long words, repeated characters, odd character sequences)
    • GT set shows no single feature is good at distinguishing between the classes
  • To do
    • account for different languages
    • consider: add layout information as input, e.g. page type

Vragen:

  • over de ground truth, waarom worden scans met annotator disagreement weggelaten? (gelaagde/getrapte benadering van annotator disagreement als alternatief voorgesteld)
  • Is het zinnig om temporele aspecten op te nemen in de ground truth? E.g. aparte ngram distributies voor begin 17de eeuw, eind 17de eeuw, etc.

Rondvraag

  • wensen voor de volgende keer: NER en entity linking/event detection
  • GLOBALISE built a transcript viewer.
    • Globalise Labs Transcriptions Viewer. https://transcriptions.globalise.huygens.knaw.nl: this is just a simple, interim tool for users, made available in conjunction with the first, full release of the 4.7M VOC transcripts. Note: it’s not the research portal or VRE for Globalise. This is also explained on the landing page.
    • https://lab.globalise.huygens.knaw.nl is the place to store all the experiments and beta versions of tools and resources from the Globalise project.