Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Show Programmes to Data: Designing a Workflow to Make Performing Arts Ephemera Accessible Through Language Models

Clarisse Bardiot, Pierre-Carl Langlais, Bernard Jacquemin, Jacob Hart, Antonios Lagarias, Nicolas Foucault, Aurélie Lemaître-Legargeant, Jeanne Fras

TL;DR

The paper tackles the challenge of making theatre programme ephemera discoverable and analysable by converting unstructured programme content into structured, interoperable data. It presents a three-stage workflow that segments documents, transcribes text with multimodal language models, and encodes results as RDF using a Linked Art–based ontology extended for performing arts, coupled with a Pleias-Wikidata reasoning model (POntAvignon). This approach yields high segmentation accuracy and robust transcription while enabling alignment with knowledge graphs, demonstrated on the Festival d'Avignon corpus. The work highlights the potential for large-scale, ontology-driven performing arts historiography and outlines practical steps toward standardised, reusable corpora, while noting reproducibility and vocabulary coverage challenges as future work.

Abstract

Many heritage institutions hold extensive collections of theatre programmes, which remain largely underused due to their complex layouts and lack of structured metadata. In this paper, we present a workflow for transforming such documents into structured data using a combination of multimodal large language models (LLMs), an ontology-based reasoning model, and a custom extension of the Linked Art framework. We show how vision-language models can accurately parse and transcribe born-digital and digitised programmes, achieving over 98% of correct extraction. To overcome the challenges of semantic annotation, we train a reasoning model (POntAvignon) using reinforcement learning with both formal and semantic rewards. This approach enables automated RDF triple generation and supports alignment with existing knowledge graphs. Through a case study based on the Festival d'Avignon corpus, we demonstrate the potential for large-scale, ontology-driven analysis of performing arts data. Our results open new possibilities for interoperable, explainable, and sustainable computational theatre historiography.

From Show Programmes to Data: Designing a Workflow to Make Performing Arts Ephemera Accessible Through Language Models

TL;DR

The paper tackles the challenge of making theatre programme ephemera discoverable and analysable by converting unstructured programme content into structured, interoperable data. It presents a three-stage workflow that segments documents, transcribes text with multimodal language models, and encodes results as RDF using a Linked Art–based ontology extended for performing arts, coupled with a Pleias-Wikidata reasoning model (POntAvignon). This approach yields high segmentation accuracy and robust transcription while enabling alignment with knowledge graphs, demonstrated on the Festival d'Avignon corpus. The work highlights the potential for large-scale, ontology-driven performing arts historiography and outlines practical steps toward standardised, reusable corpora, while noting reproducibility and vocabulary coverage challenges as future work.

Abstract

Many heritage institutions hold extensive collections of theatre programmes, which remain largely underused due to their complex layouts and lack of structured metadata. In this paper, we present a workflow for transforming such documents into structured data using a combination of multimodal large language models (LLMs), an ontology-based reasoning model, and a custom extension of the Linked Art framework. We show how vision-language models can accurately parse and transcribe born-digital and digitised programmes, achieving over 98% of correct extraction. To overcome the challenges of semantic annotation, we train a reasoning model (POntAvignon) using reinforcement learning with both formal and semantic rewards. This approach enables automated RDF triple generation and supports alignment with existing knowledge graphs. Through a case study based on the Festival d'Avignon corpus, we demonstrate the potential for large-scale, ontology-driven analysis of performing arts data. Our results open new possibilities for interoperable, explainable, and sustainable computational theatre historiography.

Paper Structure

This paper contains 13 sections, 8 figures, 5 tables.

Table of Contents

  1. From Pdf to text
  2. From programmes to pages
  3. From page to text
  4. Defining an ontology for performing arts
  5. From ontology to policy -- training a reasoning model for semantic data annotation
  6. Future Work
  7. Segmentation pretest
  8. Separator identification (segmentation pipeline -- step 2)
  9. Reference widths used for the preprocessing (segmentation)
  10. Final Claude prompt
  11. DAN prompt prefix (inserted ahead of the previous prompt)
  12. Samples of generation after RL training
  13. EasyOCR performances (pre-VLM OCR evaluation)

Figures (8)

  • Figure 1: Overview of the workflow to transform show programmes into structured data.
  • Figure 2: Examples of born digital and digitised show programmes from the Festival d’Avignon corpus. Top: J. Gosselin, 2666, Avignon 2016. Born-digital show programme; “flyer” format, with front and back pages (page 1 and page 2 respectively), each containing 3 subpages. Bottom: W. Forsythe, Ballet Frankfurt, Avignon 1991. Digitised document. Pages 1 and 5 each contain a single subpage. Other ones contain multiple subpages to be segmented. On page 3, the text is spanned across 2 subpages, and should not be segmented to preserve its original textual content.
  • Figure 3: Overview of our segmentation pipeline from programmes to pages.
  • Figure 4: Conceptual schema of the ontology for performing arts: Work (A) -- Production (B) -- Show (C).
  • Figure 5: Training cycle of the POntAvignon model, showing the evolution of GRPO reward across training steps.
  • ...and 3 more figures