Hierarchical Narrative Overview

• Hierarchical Narrative is defined as a multi-level organization of narrative content, where local units are preserved and higher-level summaries support inference and generation.
• Various methodologies—including cognitive generative models, graph-based representations, and contrastive clustering—are employed to capture and analyze the narrative hierarchy across different media.
• Applications span film-script analysis, long-form text summarization, and multimodal video understanding, demonstrating measurable improvements in narrative structure evaluation and generation.

Hierarchical narrative denotes a class of representations in which narrative content is organized across nested levels rather than treated as a flat sequence. In the cited literature, those levels may be cognitive states within a generative model, beats within scenes, scenes within acts, frames within shots and scenes, event thoughts within chapters, or scene units within storyline trees. The same idea appears in explanatory psychiatry, film-script analysis, long-form question answering, multimodal video understanding, structured summarization, and generative systems for text, music, and audio-visual media (Moore, 2015, 0805.3799, Ou et al., 18 Jun 2026, Asgarov et al., 18 Jun 2026). This suggests a recurrent design principle: local units are preserved, and higher-level nodes, summaries, or latent states are added above them to support inference, generation, or analysis.

1. Formal definitions and representational families

One influential formalization defines a formal narrative as a sequence of internal cognitive states,

$N=\{W_1,W_2,\dots,W_j\},$

with probability

$$p(N_k)=p(W_1)\,p(W_2\mid W_1)\,p(W_3\mid W_1,W_2)\dots p(W_j\mid W_1,\dots,W_{j-1}),$$

and, in practice, an $n$-gram approximation

$$p(N_k)\simeq p(W_1)\prod_{i=2}^j p(W_i\mid W_{i-1}).$$

In that framework, narratives are generated and inferred within a hierarchy of cognition, and multiple narrative-based inferences are integrated into a single conscious “view of the world” or unitary perspective (Moore, 2015).

A second family of definitions uses explicit graph structure. Hierarchical Narrative Analysis defines a rooted, directed acyclic graph $H=(V,E,\ell)$, with $\ell(n)\in\{0,1,2\}$, where level 0 contains top-level claims or requests, level 1 contains primary perceptions, and level 2 contains secondary perceptions; every edge satisfies $\ell(n_j)=\ell(n_i)+1$ (Matsuoka et al., 2024). Storyline trees define a rooted directed acyclic graph $T=(V,E)$ whose internal nodes are storyline nodes $t=(\tau,\delta)$, with $\tau$ a title and $$p(N_k)=p(W_1)\,p(W_2\mid W_1)\,p(W_3\mid W_1,W_2)\dots p(W_j\mid W_1,\dots,W_{j-1}),$$0 a one-paragraph description, and whose leaves are scenes (Ou et al., 18 Jun 2026).

A third family adopts multilevel event or knowledge graphs. In the comics framework, the hierarchy consists of panel-level multimodal graphs $$p(N_k)=p(W_1)\,p(W_2\mid W_1)\,p(W_3\mid W_1,W_2)\dots p(W_j\mid W_1,\dots,W_{j-1}),$$1, sequence-level temporal graphs $$p(N_k)=p(W_1)\,p(W_2\mid W_1)\,p(W_3\mid W_1,W_2)\dots p(W_j\mid W_1,\dots,W_{j-1}),$$2, and event-level semantic graphs $$p(N_k)=p(W_1)\,p(W_2\mid W_1)\,p(W_3\mid W_1,W_2)\dots p(W_j\mid W_1,\dots,W_{j-1}),$$3, unified by cross-level edges such as $$p(N_k)=p(W_1)\,p(W_2\mid W_1)\,p(W_3\mid W_1,W_2)\dots p(W_j\mid W_1,\dots,W_{j-1}),$$4, $$p(N_k)=p(W_1)\,p(W_2\mid W_1)\,p(W_3\mid W_1,W_2)\dots p(W_j\mid W_1,\dots,W_{j-1}),$$5, and $$p(N_k)=p(W_1)\,p(W_2\mid W_1)\,p(W_3\mid W_1,W_2)\dots p(W_j\mid W_1,\dots,W_{j-1}),$$6 (Chen, 14 Apr 2025). In long-video understanding, NEST represents each event as

$$p(N_k)=p(W_1)\,p(W_2\mid W_1)\,p(W_3\mid W_1,W_2)\dots p(W_j\mid W_1,\dots,W_{j-1}),$$7

where $$p(N_k)=p(W_1)\,p(W_2\mid W_1)\,p(W_3\mid W_1,W_2)\dots p(W_j\mid W_1,\dots,W_{j-1}),$$8 is a PropBank-grounded trigger, $$p(N_k)=p(W_1)\,p(W_2\mid W_1)\,p(W_3\mid W_1,W_2)\dots p(W_j\mid W_1,\dots,W_{j-1}),$$9 are semantic role arguments, and $n$0 is the time span; relations include $n$1, which marks parent–child composition between macro-events and micro-events (Asgarov et al., 18 Jun 2026).

2. Narrative units and hierarchical scales

Hierarchical narrative work differs sharply in its choice of basic unit, but the levels are consistently ordered from fine-grained observations to broader narrative abstractions.

Domain	Units	Source
Cognition and OCD	raw sensory features → simple percepts → proto-narratives → full-blown narratives	(Moore, 2015)
Film scripts	beats → scenes → sequences/acts	(0805.3799)
Movies	frames → shots → scenes	(Wang et al., 2023)
Comics	panels → sequences → events	(Chen, 14 Apr 2025)
E-commerce video captioning	TCoT segments → chapter summaries	(Li et al., 12 Jan 2026)
Long-form novels	scenes → storyline nodes at multiple abstraction levels	(Ou et al., 18 Jun 2026)

In the cognitive model, the lowest level contains “raw sensory features,” the next contains “simple percepts,” intermediate levels contain “mid-level percepts and proto-narratives,” and the highest levels contain “full-blown narratives about events in time” (Moore, 2015). In Murtagh et al.’s film-script analysis, the script is segmented into scenes, or into beats when analyzing a particularly important scene, and chronology-respecting clustering is used to recover beats, sequences, acts, and climaxes (0805.3799).

In long-form video, MoviePuzzle explicitly decomposes clips into frame, shot, and scene levels. Frames are aligned with subtitle utterances; shots partition frames into groups sharing a shot id; scenes partition shots into groups with the same scene id (Wang et al., 2023). HiVid-Narrator uses a different pairing: event-level “Temporal Chain-of-Thought” units,

$n$2

and higher-level chapter summaries,

$n$3

with E-HVC-146K reporting $n$4 and $n$5 (Li et al., 12 Jan 2026).

Storyline Trees argues that scenes, rather than chapters or generic segments, are the appropriate basic units because chapters are “too coarse” and fixed-length or semantic chunks “disrupt event boundaries.” Scenes are defined as contiguous narrative segments where time, location, or character configuration remain stable until a meaningful change occurs (Ou et al., 18 Jun 2026). This emphasis on scene structure recurs across film analysis, long-form QA, and video captioning.

3. Inference, composition, and structure induction

Hierarchical narrative systems differ not only in representation but also in how they infer, compose, or learn structure. In the cognitive generative account, perceptual inference at each level applies a Bayes update,

$n$6

while narrative generation samples future or alternative states using a transition model $n$7. The paper describes perception and generation as complementary “analysis-by-synthesis” loops and uses a threat-recognition example in which a Security Motivational System generates alternative threat narratives and scans them for protective action (Moore, 2015).

In film-script analysis, the workflow begins with a scenes-by-words contingency table $n$8, derives row profiles, embeds scenes in a Correspondence Analysis factor space that preserves $n$9 distances, and then applies adjacency-constrained complete-link hierarchical clustering. Because only adjacent clusters may merge, the resulting dendrogram respects chronology. Murtagh et al. report that this procedure recovers McKee’s notions of beats, sequences, acts, and climax, and validate the structure by randomizing scene order $$p(N_k)\simeq p(W_1)\prod_{i=2}^j p(W_i\mid W_{i-1}).$$0 and comparing nine style-and-structure attributes against the real script (0805.3799).

MoviePuzzle learns hierarchy directly. Its Hierarchical Contrastive Movie Clustering model uses CLIP-based vision and text encoders, a 2-layer 8-head Transformer with hidden size 512, a pairwise binary ordering head $$p(N_k)\simeq p(W_1)\prod_{i=2}^j p(W_i\mid W_{i-1}).$$1, and a contrastive clustering head $$p(N_k)\simeq p(W_1)\prod_{i=2}^j p(W_i\mid W_{i-1}).$$2. At each of the frame, shot, and scene levels it optimizes

$$p(N_k)\simeq p(W_1)\prod_{i=2}^j p(W_i\mid W_{i-1}).$$3

with total loss

$$p(N_k)\simeq p(W_1)\prod_{i=2}^j p(W_i\mid W_{i-1}).$$4

and reports that $$p(N_k)\simeq p(W_1)\prod_{i=2}^j p(W_i\mid W_{i-1}).$$5 is best. The model shares encoders across levels but uses distinct heads, so embeddings become “temporally and hierarchically aware at all scales” (Wang et al., 2023).

Graph-based approaches make composition explicit. In the comics framework, the unified graph

$$p(N_k)\simeq p(W_1)\prod_{i=2}^j p(W_i\mid W_{i-1}).$$6

supports graph traversal for action retrieval, dialogue tracing, character appearance mapping, and panel timeline reconstruction (Chen, 14 Apr 2025). NEST formalizes hierarchical composition as one relation type within a six-label event graph, with the example “leave home” decomposed into “grab keys,” “open door,” and “walk out” (Asgarov et al., 18 Jun 2026). These formulations make hierarchy operational rather than purely descriptive.

4. Text-centric planning, summarization, and revision

In long-form text generation, hierarchical narrative commonly appears as a separation between global planning and local realization. EIPE-text retains the Plan-and-Write paradigm but replaces one-shot planning with iterative plan extraction from a corpus of narratives. For each narrative $$p(N_k)\simeq p(W_1)\prod_{i=2}^j p(W_i\mid W_{i-1}).$$7, an LLM sketches an initial tree-like plan $$p(N_k)\simeq p(W_1)\prod_{i=2}^j p(W_i\mid W_{i-1}).$$8, multiple-choice QA pairs are generated to test coverage, and failed questions are converted into localized refinement instructions such as “add node …”, “modify content …”, or “adjust subtree …”. The evaluation score is

$$p(N_k)\simeq p(W_1)\prod_{i=2}^j p(W_i\mid W_{i-1}).$$9

and extraction stops when $H=(V,E,\ell)$0, improvement becomes negligible, or a maximum $H=(V,E,\ell)$1 is reached; the paper states $H=(V,E,\ell)$2 on average (You et al., 2023).

NexusSum organizes summarization into Preprocessing, Summarization, and Compression. Its Dialogue-to-Description transformation rewrites dialogue into third-person prose, and its hierarchical multi-LLM pipeline processes scene-based chunks of $H=(V,E,\ell)$3 scenes, recursively aggregating summaries across levels. Length is controlled by a target $H=(V,E,\ell)$4, a chunk size $H=(V,E,\ell)$5, and a maximum of 10 compression iterations. The paper reports a BookSum BERTScore increase from 54.4 to 70.7, corresponding to a 30.0% improvement, with additional gains on MovieSum (+7.1%) and MENSA (+1.7%), and presents a linear-time complexity analysis under fixed chunk sizes and compression ratios (Kim et al., 30 May 2025).

Dramaturge applies hierarchy to script revision rather than initial generation. Its pipeline consists of Global Review, Scene-level Review, and Hierarchical Coordinated Revision. Four global evaluator agents produce high-level suggestions, four scene-level inspectors identify local flaws under guidance $H=(V,E,\ell)$6, and revision is coordinated through a Storyline Editor, Scene Editors, Dialogue Editors, a Script Description Editor, and a final Script Polisher. The process is coarse-to-fine, top-down, and iterative. On 50 scripts, the paper reports script-level overall evaluation improving from 57.18 to 87.70 and scene-level comparative evaluation from approximately 56.5 to 94.2; against Gemini-2.5-pro, the reported margin is +8.3 points at script level and +19.9 points at scene level (Xie et al., 6 Oct 2025).

These systems converge on the same architectural claim: narrative quality depends on separating structural revision from sentence-level or scene-level realization. That claim is explicit in the divide-and-conquer workflow of Dramaturge, the plan-first factorization of EIPE-text, and the staged aggregation of NexusSum.

5. Multimodal grounding and generation

In multimodal settings, hierarchical narrative often serves as an interface between raw signals and controllable generation. HiVid-Narrator constructs the E-HVC dataset with dual-granularity annotations: temporally grounded event-level TCoT and chapter-level summaries. Its staged annotation pipeline performs ASR-driven segmentation and enhancement, frame-level description, then coarse-to-fine refinement of TCoT and chapters. To manage information density, the Scene-Primed ASR-anchored Compressor fuses ASR and vision tokens, forms scene and event representations, and assembles hierarchical tokens. With $H=(V,E,\ell)$7, $H=(V,E,\ell)$8, and $H=(V,E,\ell)$9, the paper reports an 82.6% reduction relative to the original $\ell(n)\in\{0,1,2\}$0 vision tokens. On E-HVC-Bench, HiVid-Narrator with SPA reports SODA$\ell(n)\in\{0,1,2\}$1 14.48, CIDEr 1.45, METEOR 32.01, and BERTScore 74.25 (Li et al., 12 Jan 2026).

NarraScore addresses long-video soundtrack generation by treating emotion as a compressed proxy for narrative logic. Each frame $\ell(n)\in\{0,1,2\}$2 is mapped to a Valence-Arousal vector

$\ell(n)\in\{0,1,2\}$3

predicted by a frozen instruction-tuned VLM with a small probe head. Conditioning is split between a Global Semantic Anchor, which stabilizes “style, genre, instrumentation, and pacing,” and a Token-Level Affective Adapter, which injects local tension into shallow decoder layers. On a dedicated long-video test set, NarraScore reports FAD 1.923, FD 36.411, KLD 0.320, and ImageBind 0.219; in long-form subjective evaluation, Emotional Dynamic Consistency reaches 2.86 and Long-term Coherence 3.15 (Wen et al., 9 Feb 2026).

MAVIN extends hierarchy to multi-shot audio-visual generation. It represents a narrative as a global caption $\ell(n)\in\{0,1,2\}$4, shot-level captions $\ell(n)\in\{0,1,2\}$5, and role-level descriptors $\ell(n)\in\{0,1,2\}$6, produced by a three-agent scripting pipeline consisting of a Structure Parser, Identity Aligner, and Narrative Refiner. Boundary-aware attention applies a binary routing mask so that each latent token only attends to text tokens whose interval covers its timestamp, while ID-aware propagation binds role-specific visual and audio anchors to preserve appearance and vocal timbre. On a 1,000-clip benchmark in the joint T2AV setting, MAVIN reports FVD 231.6, FAD 6.8, TVS 0.2471, TAS 0.2392, WER 0.048, Sync 6.032, TAMS 0.8104, and STA 0.9897 (Liu et al., 28 Jun 2026).

Across these systems, hierarchical narrative is not merely descriptive metadata. It becomes a control surface for compression, alignment, and conditional generation.

6. Evaluation, interpretation, and unresolved difficulties

Hierarchical narrative has also become a method for analysis and evaluation. Hierarchical Narrative Analysis uses a three-stage LLM pipeline to extract Patterns 1–4, then Pattern 5, then summarize edges into compact A–F sentences. Applied to 2,998 public comments on generative AI, it reports 5,637 normative claims, 1,892 requests, 8,527 causal relations, 6,227 perceptions, and 6,054 hierarchical links, together with a “Zero OOV Guarantee” and separate positive and negative narrative networks (Matsuoka et al., 2024). H3Prompt uses a three-step prompting hierarchy for multilingual news classification—domain, main narrative, then sub-narrative—over a taxonomy with 20 main narratives and 181 sub-narratives; on the English test set it placed first among 28 teams, and on the English development set the reported scores are 0.577 macro F1 for main narratives and 0.482 samples F1 for sub-narratives (Singh et al., 28 May 2025).

Recent benchmarks also show that long context alone does not guarantee narrative understanding. NEST contains 1005 full-length movies of average 98 minutes, each annotated with 102 multimodal narrative events and relations including temporal ordering, long-range dependencies, and hierarchical composition. The benchmark is “highly challenging for grounded event discovery,” with ETD below 8%, EL under 6%, and EAE below 11%. Event Relation Extraction is more tractable once events are given, reaching 35.45% F1 zero-shot and 44.42% F1 after fine-tuning; hierarchical F1 rises from 11.76% to 38.26% after fine-tuning Qwen3-Omni (Asgarov et al., 18 Jun 2026).

A comparable shift appears in narrative-based mental-health modeling. Across 830 Chinese therapeutic texts, the three-level framework of micro lexical features, meso semantic embeddings, and macro LLM narrative evaluation reports that macro-level evaluation “substantially outperforms lexical and embedding features.” Macro-level regression alone yields $\ell(n)\in\{0,1,2\}$7 for depression and $\ell(n)\in\{0,1,2\}$8 for anxiety, while the full model reaches $\ell(n)\in\{0,1,2\}$9 and $\ell(n_j)=\ell(n_i)+1$0; in classification, macro-level AUC is 0.647 for depression, 0.661 for anxiety, and 0.692 for trauma (Ma et al., 30 Apr 2026).

The interpretive consequence is consistent across domains. Conventional text mining methods such as sentiment analysis and topic modeling are reported as limited in their ability to capture hierarchical narrative structures (Matsuoka et al., 2024). Long-context models can process extended token streams yet still fail on the interaction of low-level actions, events, and long-range narrative dependencies (Asgarov et al., 18 Jun 2026). In the OCD framework, a wide variety of symptoms is explained by “a single dysfunction” in sub-surface levels of inference while the global unitary perspective remains intact (Moore, 2015). A plausible implication is that future progress will depend less on larger undifferentiated context windows and more on explicit intermediate structure—scenes, events, plans, graphs, captions, or affective trajectories—that can be independently grounded, revised, and evaluated.