IA et description des archives numériques

Could we have item-level descriptions generated by AI and then build synthetic descriptions for archival files?

This preliminary study presents one use of artificial intelligence for the description of archival documents. This exploratory study is based on a small sample of documents showing slight variations, making it possible to more precisely assess the impact of different variables such as typological complexity, number of dates, file formats, presence of machine readable structures, and amount of named entities.

Of particular interest is the emphasis on the methodology used to formalize variations of the same document and the preliminary results obtained. The approach relies on a factorial experiment, with result interpretation based on a scoring system that takes into account relevance, stability and ability to process documents.

For this first try, the free version of Gemini, Copilot, ChatGPT and Claude has been used, along with a locally deployed instance of Qwen. This initial study not only highlights the context of certain cases that are easier or more difficult for AI to characterize, but also reveals systematic patterns in AI performance, helping to identify strengths, limitations, and directions for more targeted and specialized future experiments, while providing insights into otherwise opaque AI decision-making processes.