Scholarlink

GLAM research often involves the collection of unstructured textual data which is semantically rich, but labour-intensive to handle and process. This paper explores the use of Large Language Models (LLMs) to support the transformation of such material into structured data that can be queried and quantitatively analysed. Focusing on a corpus of notes documenting the closure of over 500 UK museums between 2000 and 2025, we present a two-stage pipeline to automate the generation of data models. In the first stage, an LLM proposes schema fragments based on chunks of notes; in the second, the LLM collates these fragments into a coherent data model. As a preliminary evaluation, we introduce a method based on syntactic validity to assess the structure of the generated models in the absence of ground truth. Our experiments with Llama 3.1 show promising results using zero-shot prompting, though ensuring semantic consistency and model integration remain challenging.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

From Notes to Models: Leveraging LLMs for Museum Closure Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • George A. Wright,
  • Andrea Ballatore,
  • Alexandra Poulovassilis,
  • Peter T. Wood

摘要

GLAM research often involves the collection of unstructured textual data which is semantically rich, but labour-intensive to handle and process. This paper explores the use of Large Language Models (LLMs) to support the transformation of such material into structured data that can be queried and quantitatively analysed. Focusing on a corpus of notes documenting the closure of over 500 UK museums between 2000 and 2025, we present a two-stage pipeline to automate the generation of data models. In the first stage, an LLM proposes schema fragments based on chunks of notes; in the second, the LLM collates these fragments into a coherent data model. As a preliminary evaluation, we introduce a method based on syntactic validity to assess the structure of the generated models in the absence of ground truth. Our experiments with Llama 3.1 show promising results using zero-shot prompting, though ensuring semantic consistency and model integration remain challenging.