From psalms to patterns:

large-scale detection of liturgical structures in Books of Hours through text reuse detection

Svetlana Yatsyk – IRHT – DH2025

Biblissima Logo

Introduction

Project Overview

HORAE Project

Cross-disciplinary Nature

  • Computer Vision
  • Natural Language Processing
  • Book History

Partners

  • IRHT (CNRS) – France
    Dominique Stutzmann, Louis Chevalier, Laura Lebarbey, Jacob Currie
  • Teklia – France
    Christopher Kermorvant, Marie-Laurence Bonhomme, Mélanie Boillet
  • LS2N – Nantes
    Béatrice Daille, Amir Hazem

Canonical Hours

  • Matins – at midnight
  • Lauds – in the early morning
  • Prime – first hour of the daylight
  • Terce – at 9 AM
  • Sext – at noon
  • None – at 3 PM
  • Vespers – at sunset
  • Compline (Completorium) – at the end of the day.

Books of Hours


  • Late medieval bestsellers (>10 000 copies)
  • For religious practices of rich lay people
  • Highly structured
  • Lavishly illuminated.

Structure of the Books of Hours

  • Liturgical calendar
  • Four gospel pericopes
  • Hours of the Virgin (Officium parvum beatae Mariae Virginis)
  • Hours of the Cross
  • Hours of the Holy Spirit
  • The office of the dead (Officium Defunctorum)
  • Penitential psalms
  • Hours of the passion
  • Suffrages
  • The litany of the saints…

Book of Hours: granularity

  • Book of Hours
    • Calendar
    • Gospel lessons
    • Hours of the Virgin
      • Matins
        • Invitatory
          • Versicles & Responses
            • Vers.
            • Resp.
          • Gloria
          • Psalm 94
            • Antiphon
            • Psalm
              • Ps 94:1-2
              • Antiphon
              • Ps 94:3-4
              • Antiphon part 2
              • Ps 94:5-7
              • Antiphon
              • Ps 94:8-9
              • Antiphon part 2
        • Hymn
        • Nocturns
          • Nocturn 1
            • Psalms
              • Psalm 1
              • Psalm 2
              • etc
            • Versicles & Responses
            • Lord’s Prayer
            • Absolution
            • Lessons
              • Lesson 1
                • Blessing
                • Lesson 1
                • Responsory 1
                  • Respond
                  • Verse
                  • Partial respond
                  • Lesser doxology
              • Lesson 2
              • Etc.
          • Nocturn 2
          • Nocturn 3
        • Canticle
      • Lauds
      • Prime
      • Terce
      • Sext
      • None
      • Vespers
      • Compline
    • Hours of the Cross & Holy Spirit
    • Penitential Psalms and Litany
    • Office of the Dead
    • Obsecro te, O intemerata
    • Suffrages
    • Accessory texts

Source:
Dominique Stutzmann, Jacob Currie, Béatrice Daille, Amir Hazem, Christopher Kermorvant. Integrated DH. Rationale of the HORAE Research Project. Digital Humanities Conference, Jul 2019, Utrecht, Netherlands.

Goals

Using text reuse detection tools, find psalms and other texts in the Books of Hours in order to:

  • describe their contents and improve the cataloguing
  • study manuscripts’ circulation and origins.

Corpus overview

  • 1 113 manuscripts, 45 early printed books
  • 1 116 HTR-sed
  • 302 772 pages
  • 13th to 16th centuries
  • available on Arkindex.
    It also holds:
  • 24 830 miniatures
  • 1 778 552 initials.

1. Data collection and processing

  • Image collection via the IIIF protocol
  • Image segmentation: detection of initials, miniatures, and text zones
  • Automatic transcription by TEKLIA using a Kaldi-based model.
HTR and Segmentation Example

Source:
Béatrice Daille, Amir Hazem, Christopher Kermorvant, Martin Maarand, Marie-Laurence Bonhomme, et al.
Transcription automatique et segmentation thématique de livres d'heures manuscrits.
Revue TAL : traitement automatique des langues, 2019, TAL et humanités numériques, 60(3), pp.13–36. (updated)

Entraînement Validation Test
WER 14.51 24.46 34.19
WER (rescored) – 32.36 26.32
CER 8.93 11.07 11.21
CER (rescored) – 14.10 9.90

2. Text reuse detection with passim

Challenges

  • Noisy HTR data
  • Little to no ground truth available for training or evaluation
  • Large amount of data to process (by medievalist standards).

Text reuse detection with passim

  • An open-source tool developed by David A. Smith.
  • Project repository: github.com/dasmiq/passim
  • Introduction and tutorial: Matteo Romanello & Simon Hengchen, Detecting Text Reuse with Passim, Programming Historian, 2021
  • Best evaluation results on our data:
    Precision: 0.76, recall: 0.51, F1 Score: 0.61

Passim output in jsonl

        "uid": -7747811833890938386,
        "cluster": 1, "begin": 231, "end": 379, "id": 398134, "series": "psalms",
        "text": "animam meam humiliavit in terra vitam meam. Collocavit me in obscuris sicut mortuos seculi et anxiatus est super me spiritus meus in me turbatum es"

        ---

        "uid": -1932149852289013984, "cluster": 1, "begin": 0, "end": 153,
        "src": [{"uid": -7747811833890938386, "begin": 231, "end": 379}],
        "id": 202,
        "series": "États-Unis, Baltimore, The Walters Art Museum, W.0085_HTR",
        "text": "animam meam humi miliavit in terra vitam meam Tollocavit me Filii obscuris sicut mortuos seculi et anxiatus est super me spiritus meus in me turbatum est"
Passim output for Paris, BNF, Ms. lat. 1406

3. Psalm sequence extraction and validation

Prediction validation

Sequences detected in France, Paris, Bibliothèque nationale de France, Manuscrits, lat. 01177
Psalm 97 detected by error

Manuscript classification by use?

Psalm clusters shared by uses

Open interactive file

Next steps

  • Short and very short text detection
  • Distinguishing between complete psalms and their fragments (antiphons etc)

Bibliography

  • Mélodie Boillet, Marie-Laurence Bonhomme, Dominique Stutzmann, and Christopher Kermorvant. HORAE: an annotated dataset of books of hours. Proceedings of The 5th International Workshop on Historical Document Imaging and Processing, Sydney, NSW, Australia, September 20–21, 2019 (HIP’19).
  • Béatrice Daille, Amir Hazem, Christopher Kermorvant, Martin Maarand, Marie-Laurence Bonhomme, et al. Transcription automatique et segmentation thématique de livres d'heures manuscrits. Revue TAL : traitement automatique des langues, 2019, TAL et humanités numériques, 60(3), pp. 13–36.
  • Dominique Stutzmann, Jacob Currie, Béatrice Daille, Amir Hazem, Christopher Kermorvant. Integrated DH. Rationale of the HORAE Research Project. Digital Humanities Conference, Jul 2019, Utrecht, Netherlands.

Thank you!

svetlana.yatsyk@cnrs.fr