💻 computer science

S^2tory: Story Spine Distillation for Movie Script Summarization

The paper introduces S^2tory, a narratology-grounded framework that leverages character development trajectories and a theory-constrained reasoning agent to distill essential plot nuclei from non-linear movie scripts, achieving state-of-the-art summarization performance with high semantic fidelity.

Original authors: Mingzhe Lu, Yanbing Liu, Qihao Wang, Jiarui Zhang, Jiayue Wu, Yue Hu, Yunpeng Li, Yangyan Xu

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mingzhe Lu, Yanbing Liu, Qihao Wang, Jiarui Zhang, Jiayue Wu, Yue Hu, Yunpeng Li, Yangyan Xu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a massive, 100-page movie script. It's full of dialogue, descriptions of sunsets, characters making coffee, and long conversations about the weather. If you asked a standard computer to summarize this, it might just grab the first few sentences of every scene or pick out the longest speeches. The result? A summary that feels like a list of random facts, missing the actual story of why the movie matters.

The paper "S2tory" introduces a new way to solve this by teaching computers to think like a film director or a literary critic, rather than just a word-counting machine. Here is how it works, broken down into simple concepts:

1. The Problem: The "Noise" vs. The "Signal"

Think of a movie script like a garden.

The Satellites (Noise): These are the flowers, the decorative rocks, and the butterflies. They make the garden look nice and set the mood, but if you removed them, the garden would still stand. In a script, these are the atmospheric scenes, small talk, or emotional moments that don't actually change the plot.
The Nuclei (Signal): These are the roots and the trunk. If you cut these out, the whole tree falls over. In a script, these are the "Plot Nuclei"—the specific events that force a character to change their mind, their goal, or their situation.

Old summarization tools tried to pick the "biggest" flowers (the most important-sounding words). S2tory realizes that to tell the story, you need to find the roots.

2. The Solution: The "Story Detective" (NEAgent)

The authors created a special AI agent called NEAgent (Narrative Expert Agent). You can think of this agent as a super-smart film student who has read every theory book on storytelling.

Instead of just reading the words, NEAgent watches the characters like a hawk. It asks a very specific question for every single scene:

"If we delete this scene, does the character's journey break?"

Scenario A: A character says, "I hate you," and then walks away. If you remove this, the character never leaves, and the next scene doesn't happen. Verdict: This is a Nucleus (Keep it!).
Scenario B: Two characters are walking down a road and see a bird. If you remove this, the characters still walk down the road and meet the villain later. Verdict: This is a Satellite (Cut it!).

3. The Magic Trick: Teaching a Small Student

NEAgent is very smart, but it's also slow and expensive to run (like a PhD professor who takes days to grade one paper). You can't use it to summarize thousands of movies instantly.

So, the authors used a technique called Distillation.

Imagine the PhD professor (NEAgent) grading a stack of scripts and writing down why they kept or cut certain scenes.
They then take those notes and teach a smart high school student (a smaller, faster AI model) how to think the same way.
The student learns the logic of the professor, not just the answers. Now, the student can quickly scan a script, identify the "roots" (nuclei), and ignore the "flowers" (satellites).

4. The Result: A Tight, Rhythmic Summary

Once the student model identifies the essential "nuclei," it feeds them to a final summarizer. The result isn't just a shorter version of the text; it's a version that keeps the rhythm of the movie.

The paper tested this on the MovieSum dataset (a huge collection of movie scripts).

The Compression: They managed to shrink the scripts by about 3.5 times (keeping only ~28% of the original text).
The Quality: Despite being so short, the summaries were better than any previous method at capturing the meaning and flow of the story.
The Proof: When they tested this on book summaries (which look very different from scripts), it still worked. This proves the AI learned the logic of storytelling, not just how to read movie scripts.

In a Nutshell

S2tory is a system that teaches AI to stop counting words and start tracking character growth. By filtering out the "fluff" and keeping only the events that actually move the story forward, it creates summaries that feel like the real movie, just told much faster. It's the difference between reading a grocery list of a movie and reading a compelling synopsis that makes you want to watch it.

Technical Summary: S2tory – Story Spine Distillation for Movie Script Summarization

1. Problem Definition

Automatic summarization of movie scripts presents a unique challenge distinct from traditional text summarization. Scripts possess a non-linear, cross-cut narrative structure where surface-level saliency methods (e.g., lexical frequency or shallow structural patterns) fail to preserve the core story progression. Existing approaches, such as MENSA (saliency-based scene selection) and DiscoGraMS (character/dialogue graph modeling), rely heavily on data-driven patterns and often miss the underlying symbolic logic that governs narrative causality. Consequently, current models struggle to maintain the "narrative backbone" of long-form stories, often conflating atmospheric details with essential plot points.

2. Methodology: The S2tory Framework

The authors propose S2tory (Story Spine Distillation), a framework grounded in narratological theory (specifically Barthes' distinction between nuclei and satellites) to reconstruct the narrative logic of a script. The methodology operates in three distinct stages:

2.1 Theoretical Formulation

The framework models a story world as a quadruple $(C, E, S, R)$ , representing characters, events, states, and state transitions.

Nuclei vs. Satellites: A nucleus is defined as an event essential to the narrative trajectory; its removal breaks the continuity of character development. A satellite is an auxiliary event that enriches atmosphere or emotion but does not alter the narrative path.
Character Dynamics: The system tracks character state transitions ( $r: s_t \to s_{t+1}$ ) induced by events. Transitions are classified as intrinsic (internal reflection) or extrinsic (external triggers). An event is deemed a nucleus if its removal prevents a detectable state change in a character's attributes (goals, identity, affiliations).

2.2 Narrative Expert Agent (NEAgent)

To operationalize this theory, the authors introduce the NEAgent, an In-Context Learning (ICL) agent powered by GPT-4o.

Function: The NEAgent processes the screenplay scene-by-scene, maintaining a rolling memory of evolving character states and event dependencies.
Reasoning: It performs counterfactual reasoning on each narrative unit (sentence/line). It evaluates whether removing a specific unit would disrupt the structural continuity of character trajectories.
Output: The agent classifies units as either nuclei ( $\kappa=1$ ) or satellites ( $\kappa=0$ ), effectively distilling the "story spine."

2.3 Distillation and Summarization Pipeline

Recognizing that the NEAgent is computationally expensive for large-scale inference, the authors employ a two-step distillation process:

Reasoning Distillation: A student model (Qwen2.5-7B) is fine-tuned using a dataset ( $D_{distill}$ ) containing naturalized narrative text, the NEAgent's symbolic reasoning traces, and the identified nuclei sets. This teaches the smaller model to mimic the expert's logic in identifying plot nuclei.
Nuclei-Conditioned Summarization: A final summarizer (Qwen2.5-0.5B) is trained on a dataset ( $D_{sum}$ ) where the input is the distilled set of nuclei ( $B_i$ ) and the target is the reference summary. This model generates the final summary conditioned strictly on the identified plot nuclei.

3. Key Contributions

Narratologically-Grounded Framework: The introduction of S2tory, which explicitly models narrative structure through character-arc trajectories rather than relying solely on surface-level text patterns.
Expert Agent and Distillation: The design of the NEAgent for theory-constrained reasoning and a novel distillation pipeline that transfers this complex reasoning capability to smaller, efficient models.
State-of-the-Art Performance: Achieving superior results on the MovieSum benchmark while demonstrating strong zero-shot generalization to the BookSum dataset (prose), validating that the model learns narrative logic rather than screenplay-specific formatting.

4. Experimental Results

Experiments were conducted on the MovieSum dataset (screenplays) and the BookSum dataset (prose).

Performance on MovieSum: S2tory achieved state-of-the-art results, with ROUGE-1 of 45.98 and ROUGE-L of 42.45. It significantly outperformed extractive baselines (e.g., TextRank, Lead-1024) by 32–38% and matched the performance of abstractive models while operating at a much higher compression ratio (28.4% vs. 55.2% for others).
Semantic Fidelity: The model achieved the highest BERTScore Recall (59.36) among all methods, indicating superior alignment with reference summaries.
Ablation Study: Removing the character trajectory profiling module from the NEAgent caused a significant drop in BERTScore F1 (from 59.23 to 53.69), confirming that modeling character development is critical for identifying narrative nuclei.
Human Evaluation: Human raters (trained in narrative theory) scored S2tory highly on indispensability (3.84/5), coherence (3.91/5), and character consistency (4.18/5). Notably, human scores for "Satellite Reduction" were higher than automated metrics, suggesting the model successfully filters non-essential content better than standard LLMs.
Out-of-Domain Generalization: When applied to BookSum without fine-tuning, over 85% of the generated nuclei were judged as narratively essential by LLM-as-Judge evaluators, demonstrating robust generalization beyond screenplay formatting.

5. Significance and Claims

The paper argues that effective long-form summarization, particularly for complex narratives like movie scripts, requires a shift from data-driven pattern matching to symbolic representation combined with neural abstraction.

Theoretical Validity: The authors claim that narratological theory provides an "indispensable foundation" for modeling non-linear narratives. By explicitly distinguishing between structural necessity (nuclei) and surface prominence (satellites), the framework overcomes the limitations of saliency-based approaches.
Efficiency vs. Quality: S2tory demonstrates that it is possible to achieve high-fidelity summarization with significantly higher compression ratios (approx. 3.5×) compared to standard abstractive methods, without sacrificing semantic coherence.
Generalizability: The success on BookSum suggests that the core capability of S2tory is the internalization of character-centric causal reasoning, allowing it to generalize across different narrative formats (scripts vs. prose) rather than overfitting to specific dataset structures.

In conclusion, S2tory posits that recovering the "story spine" through character development trajectories is the key to unlocking high-quality summarization for structurally complex, long-form narratives.