💬 NLP

Storyline Trees: Hierarchical Representations for Long-Form Narratives

This paper introduces "storyline trees," a hierarchical representation that organizes long-form narratives into multi-level structures of scenes and plotlines to enable adaptive retrieval, which significantly improves question-answering performance compared to existing long-context models and chunk-based methods.

Original authors: Litu Ou, Mirella Lapata

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Litu Ou, Mirella Lapata

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 are trying to solve a mystery, but instead of a few clues on a table, you are handed a library containing 500 novels. You need to find one specific detail: Why did the character leave town in Chapter 42?

If you just dump all 500 books into a giant pile and ask a super-smart robot (an AI) to read them all at once, the robot gets overwhelmed. It's like trying to find a single needle in a haystack that is the size of a mountain. Even if the robot is very smart, it often gets "lost in the middle," forgetting the beginning or missing the crucial clue hidden deep in the text.

This paper introduces a new way to organize those books so the robot doesn't get lost. They call it Storyline Trees.

The Problem: The "Flat" Pile

Currently, most AI models treat a long book like a flat list of words. They don't really understand that a story has a structure. They see Chapter 1, then Chapter 2, then Chapter 3, but they don't necessarily understand that Chapter 10 is actually the climax of a plotline that started in Chapter 2.

The authors say, "Let's stop treating books like a flat wall of text and start treating them like a tree."

The Solution: Building a Storyline Tree

Think of a long novel not as a single long road, but as a giant family tree.

The Leaves (The Scenes):
First, the authors break the book down into its smallest, most logical chunks called scenes.
- Analogy: Imagine a movie. A "scene" is when the characters are in one room, doing one thing, at one time. If the characters move to a different room or the time jumps forward, that's a new scene.
- The authors use AI to cut the book into these scenes automatically. This is much better than just cutting by "Chapter," because a chapter might contain three different scenes, or a single scene might spill over two chapters. Scenes are the true building blocks of a story.
The Trunk and Branches (The Plotlines):
Once they have all the scenes, they build the tree in two ways:
- Top-Down (The Architect): They look at the whole book and ask, "What are the big themes?" (e.g., "Heathcliff's Revenge"). Then they break that big theme down into smaller sub-themes, and those into specific scenes. It's like an architect drawing a blueprint from the roof down to the foundation.
- Bottom-Up (The Builder): They start with the individual scenes and ask, "Which of these scenes belong together?" They group similar scenes (like all the scenes where the hero is sad) and summarize them into a bigger "branch." They keep doing this until they reach the top of the tree.

How the Robot Uses the Tree (Adaptive Retrieval)

This is the magic part. When you ask the AI a question, it doesn't just guess or read the whole book again. It uses the tree as a map.

The Old Way: The AI tries to guess which pages are important and reads them. If it guesses wrong, it misses the answer.
The New Way (Adaptive Retrieval):
1. The AI looks at the top of the tree (the big themes). It asks, "Which branch of the tree is likely to have the answer?"
2. It picks that branch and looks at the smaller branches underneath it.
3. It keeps zooming in, moving from the big picture down to the specific scene (the leaf) where the answer is hiding.
4. It reads only that specific scene to answer the question.

The Analogy: Imagine you are looking for a specific address in a city.

Old Method: You drive down every single street in the city until you find it.
Storyline Tree Method: You look at the city map. You see the neighborhood (Top Level), then the specific street (Middle Level), then the house number (Scene Level). You drive directly to the house.

What They Found

The researchers tested this on three different sets of long books and difficult questions. Here is what happened:

It Works Better: The "Storyline Tree" method beat every other method they tried, including models that were specifically trained to read long texts.
Scenes Matter: Breaking the book into "scenes" was much more important than just breaking it into "chapters." Chapters are often arbitrary (based on page count), but scenes are based on the actual story flow.
The Top-Down Approach Wins: While both building the tree from the top and from the bottom worked, building it from the top (starting with the big themes) worked slightly better. It gave the AI a clearer "big picture" to navigate from.

The Bottom Line

This paper shows that to help AI understand long stories, we shouldn't just give it more memory. We need to give it a structure. By organizing a book into a tree of scenes and plotlines, we give the AI a map to navigate the story, allowing it to find answers quickly and accurately without getting lost in the thousands of pages.

Technical Summary: Storyline Trees for Long-Form Narratives

Problem Statement

Long-form narratives, such as novels, present a significant challenge for current long-context language models. Unlike technical documents with explicit structural cues (headings, sections), narratives rely on implicit connections between events, characters, and plotlines distributed across hundreds of pages. Even when the entire text fits within a model's context window, performance degrades on tasks requiring the integration of dispersed information. Existing methods often fail because they either rely on explicit document structures that do not exist in fiction or treat narratives as flat collections of retrievable passages, lacking a reusable representation of the narrative's implicit structure. Furthermore, chapter boundaries are often dictated by pacing or style rather than thematic coherence, making them inconsistent units for tracing narrative threads.

Methodology

The authors propose Storyline Trees, a hierarchical representation that organizes long-form narratives from global themes and major plotlines down to fine-grained events. The methodology consists of three core components:

1. Scene Segmentation as Basic Units

Instead of using chapters or fixed-length segments, the system segments narratives into scenes. A scene is defined as a contiguous narrative unit characterized by relative consistency in time, location, character configuration, and action.

Process: Chapters are first extracted from metadata. An LLM then segments each chapter into scenes, generating structured descriptions for each: a Title (concise label), Synopsis (content summary), and Significance (narrative function).
Rationale: Scenes provide a consistent level of granularity aligned with narrative theory, avoiding the variability of chapter lengths and the continuity breaks caused by fixed-length segmentation.

2. Tree Induction

The paper introduces two complementary procedures to construct the hierarchy:

Top-Down Induction: This approach infers abstract storylines from the full set of scenes and recursively refines them into specific sub-storylines. It aligns with a "plot-first" view, where events derive meaning from their role in larger arcs. The process conditions on the full scene sequence and the previous level of storylines to ensure grounded, specific elaboration.
Bottom-Up Induction: Inspired by RAPTOR, this approach starts with scene representations, recursively clusters semantically related scenes, and summarizes each cluster into a higher-level storyline node. This aligns with a "scene-first" view, where broader storylines emerge from local relations (shared characters, causal links).

3. Adaptive Retrieval for QA

The constructed trees serve as an interface for an adaptive retrieval procedure during question answering.

Mechanism: Given a question, the model retrieves an initial set of scenes. It then iteratively inspects the storyline tree to decide which parts of the narrative to examine next.
Decision Loop: At each step, the model either commits to an answer or requests evidence from a specific storyline node. If a node is selected, its title is used as a query to retrieve the top- $k$ most relevant scenes from the original scene pool.
Advantage: This allows the model to progressively narrow its focus from broad narrative structure to localized scene evidence, guided by an interpretable map of the story rather than a flat retrieval policy.

Key Contributions

Storyline Trees: An interpretable hierarchical representation connecting global themes with fine-grained scene-level evidence, specifically designed for the implicit structure of long-form narratives.
Dual Construction Methods: The introduction of both top-down and bottom-up induction techniques, alongside an adaptive retrieval procedure that enables progressive exploration of the tree.
Empirical Validation: Demonstration of performance improvements on three long-context narrative QA benchmarks (DetectiveQA, NovelQA, LiteraryQA), with ablation studies confirming the efficacy of scene-based units and adaptive traversal.

Experimental Results

Experiments were conducted on three datasets using Qwen3 models (30B and 235B parameters) as backbones.

Performance: Top-down adaptive retrieval consistently achieved the best accuracy across all datasets and model scales, outperforming zero-shot baselines, memory-based agents (MemAgent), multi-agent tree approaches (TOA), and static retrieval methods.
Construction Direction: Top-down construction proved superior to bottom-up, suggesting that imposing a global narrative structure first and refining it downward provides a better retrieval interface than aggregating locally.
Basic Units: Scene-based construction significantly outperformed alternatives using chapters, TextTiling segments, or semantic chunks, even when retrieval budgets were matched by token count. This indicates that author-defined boundaries and generic semantic coherence are insufficient for narrative QA.
Retrieval Efficiency: The gains were not merely due to retrieving more context. Static retrieval with larger budgets did not reliably improve performance. The adaptive tree-guided approach was most effective because it actively used the tree structure to navigate from incomplete evidence to crucial scene-level details.
Question Types: The method yielded the largest gains on questions requiring the location and interpretation of narrative events within broader contexts (e.g., plot, character, setting, and meaning questions).

Significance and Claims

The paper claims that making narrative structure explicit offers a viable path toward long-context understanding that goes beyond simply extending context windows. The results suggest that:

Structure Matters: Modeling implicit storyline structure is more beneficial than treating narratives as flat collections of passages.
Adaptivity is Key: Adaptive retrieval guided by a hierarchical representation allows models to iteratively inspect high-level structure and retrieve scene-level evidence on demand, outperforming fixed or agentic chunk-based methods.
Scenes are Fundamental: Scenes serve as a more effective basic unit for narrative representation than chapters or generic segments because they align with the event structure of long-form narratives.

The authors acknowledge limitations, noting that the current system is entirely prompt-based, which may limit the reliability of traversal decisions compared to trained policies. They also note that the computational cost of tree construction is non-trivial, though it can be amortized across multiple queries on the same book. Future work is suggested to focus on learning retrieval and traversal policies directly over storyline trees and extending the representation to other narrative genres.