S^2tory: AI Framework for Movie Script Summarization Using Narrative Theory
A new framework for automatic movie script summarization, named S^2tory (Story Spine Distillation), has been developed by researchers to tackle the complexities of non-linear narratives. This innovative approach employs techniques rooted in narratology, focusing on character development arcs to pinpoint crucial plot nuclei—key events that propel the narrative—while discarding less important satellite events. A Narrative Expert Agent (NEAgent) utilizes theory-constrained reasoning to condense knowledge into a compact model for identifying these plot nuclei, which are subsequently used by another model to create summaries. Testing on the MovieSum dataset resulted in state-of-the-art semantic fidelity with about 3.5x compression, while zero-shot evaluation on BookSum demonstrated impressive out-of-domain generalization. The findings were published on arXiv under identifier 2605.03244.
Key facts
- S^2tory (Story Spine Distillation) is a framework for movie script summarization.
- It uses narratology-grounded techniques to identify plot nuclei.
- Character development trajectories are leveraged to find essential events.
- The Narrative Expert Agent (NEAgent) performs theory-constrained reasoning.
- A small model is conditioned to identify plot nuclei using distilled knowledge.
- Another model generates summaries from plot nuclei.
- Experiments on MovieSum dataset achieved state-of-the-art semantic fidelity.
- Compression ratio is approximately 3.5x.
- Zero-shot evaluation on BookSum confirmed strong out-of-domain generalization.
- Published on arXiv with identifier 2605.03244.
Entities
Institutions
- arXiv