AI-Generated Oral History Visualization: Evaluating Narrative Loss in Diaspora Interviews
A new preprint on arXiv (2607.24756) dives into how we can turn oral-history interviews from the diaspora into visual stories using AI. It points out the challenges of balancing scene planning with the story's authenticity. The study compares two approaches: a Multi-Agent Scene-decomposition pipeline (MAS) and a Single Summarization Pipeline (SSP), analyzing 82 interviews through 15 metrics based on oral-history theory to identify three failure types. The research highlights that the original narrative strength is crucial to these conflicts. The authors propose a framework to assess these failure modes and suggest a method for choosing the right system. It also discusses the changes that happen when personal memories shift into media representations.
Key facts
- arXiv:2607.24756v1
- Study compares MAS and SSP pipelines
- 82 interviews from diaspora communities
- 15 metrics around three failure modes
- Narrative-structure strength is primary predictor of conflict
- Proposes failure-mode-based evaluation framework
- Proposes routing protocol for system selection
- Focuses on double transformation of oral-history interviews
Entities
Institutions
- arXiv