About Archimedia :

Vast collections of digital multimedia content are possessed by various institutions, organizations, and companies, e.g., TV channels, radio stations, newspapers. These collections, shaped by editorial policies, contain a wide variety of document types and are rich in content and relationships. The use of deep large language and multimodal models has enabled the extraction of high-quality knowledge from these collections. However, many research problems remain open, particularly in high-stakes professional contexts where accurate and reliable results are crucial.

The ARCHIMEDIA project aims to study the scientific problems that are raised when using computers to automatically analyze these collections. Most of the problems we face are related to the impact of the properties and usage of large collections of multimedia material in professional settings onto algorithms. Some of the key problems identified include:

1. Targeted Analysis : Professionals require precise answers from collections, unlike general-purpose AI systems that provide average results.
2. Dynamic Collections : Collections constantly evolve, requiring continuous learning and updates, unlike static AI systems.
3. Long-Term Efforts : Collections are long-term efforts with evolving annotations and metadata, often poorly specified, limiting the effectiveness of AI systems.
4. Temporal Scope : Analyzing collections requires considering the temporal context of documents, as their meaning and relevance can change over time.
5. Collaborative Analysis : Collections facilitate long-term collaborative work among multiple people, but current models fail to support and track these collaborations.
6. Constrained Resources : Many AI solutions are resource-intensive and inaccessible to most organizations.

ARCHIMEDIA is a joint team with the company Ouest-France. The preferred professional context is therefore that of journalism, given our long-standing collaboration.