Track 04 – Acquisition and Digitization

Permanent URI for this collection

https://diglib.eg.org/handle/10.2312/3607221

Browse

Browsing Track 04 – Acquisition and Digitization by Subject "Artificial Intelligence → Machine learning"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    ArTLLaMA: Adaptating LLaMA to Performative Art Applications
    (The Eurographics Association, 2025) Passone, Elisa; Borazio, Federico; Hromei, Claudiu Daniel; Croce, Danilo; Basili, Roberto; Campana, Stefano; Ferdani, Daniele; Graf, Holger; Guidi, Gabriele; Hegarty, Zackary; Pescarin, Sofia; Remondino, Fabio
    Performative Arts represent a compelling and underexplored domain for the application of Generative AI, given their rich conceptual complexity and cultural depth. This paper presents ArTLLaMA, a domain-adapted version of the LLaMA language model, designed to support natural language querying of ArTBase, the first national database of Italian theatres and theatre archives. We focus on the Text-to-SQL task: automatically translating user questions into executable SQL queries. Off-the-shelf models often fail in this setting due to a lack of domain knowledge and schema awareness. To bridge this gap, we propose a two-stage fine-tuning methodology: first, we train the model to internalize the Entity-Relationship (ER) schema of ArTBase; then, we fine-tune it on a curated set of over 800 natural language-SQL query pairs reflecting real use cases in the domain. Our results show that schema-informed fine-tuning significantly boosts accuracy, with the best model achieving over 70% exact match andgenerating correct SQL even for complex queries involving multi-table joins and aggregations. Compared to general purpose models like ChatGPT, our approach yields more accurate, schema-compliant outputs. Beyond technical improvements, this work underscores the value of interdisciplinary collaboration: by embedding domain knowledge from the humanities into AI systems, we enable new forms of access, interaction, and understanding of cultural heritage data.