Hierarchical Neural Story Generation

• Hierarchical neural story generation is an approach that divides narrative creation into multi-level tasks such as plot planning, outline generation, and text realization.
• It leverages methodologies like outline-first, latent planning, and coarse-to-fine frameworks integrated with architectures such as CNNs, RNNs, and Transformers.
• Empirical evaluations show these models reduce perplexity and improve narrative coherence and consistency over traditional flat language models.

Hierarchical neural story generation is an approach to automatic narrative creation in which the generation process is decomposed into multiple levels, each corresponding to different narrative granularity—such as plot planning, outline creation, or surface realization. This architectural stratification is motivated by the observation that uninformed left-to-right language modeling struggles with long-range coherence, event consistency, and global plot structure. Hierarchical models address these deficiencies by explicitly modeling high-level abstractions (e.g., outlines, event chains, predicate-argument structures) separately from lower-level language realization, allowing more explicit control over narrative progression, character arcs, and thematic consistency.

1. Hierarchical Decomposition Paradigms

Two-stage and multi-stage decompositions are the canonical structures:

• Outline-first models generate a high-level summary or sequence of plot points and then expand each point into full text. The decomposition is formalized as

$P(Y \mid X) = \sum_Z P(Z \mid X) P(Y \mid X, Z)$

where $X$ is a prompt, $Z$ is an outline, and $Y$ is the full story (Drissi et al., 2018, Wang et al., 2020).

• Latent planning models posit a discrete, sequence-level latent variable $z = (z_1, ..., z_K)$ (e.g., anchor words, events, or topics). The generative joint is

$$p(z, x \mid t) = p_{\phi}(z \mid t) p_{\theta}(x \mid z, t)$$

with training via amortized variational inference or discrete latent variable optimization (Jhamtani et al., 2020).

• Coarse-to-fine models introduce further intermediate levels, for example generating predicate–argument structures before surface realization and then filling entity placeholders (Fan et al., 2019).

State-of-the-art hierarchical systems implement these abstractions with CNNs, RNNs, or Transformer-based architectures depending on the sub-task (Fan et al., 2018, Wang et al., 2020, Jhamtani et al., 2020).

2. Architectural Variants and Attention Mechanisms

Hierarchical neural story generators employ a variety of operational modules:

• Outline and plan generators often use convolutional sequence-to-sequence networks with self-attention (Drissi et al., 2018), Transformer decoders (Wang et al., 2020), or RNN-based planners (Jhamtani et al., 2020).
• Story or surface realization modules condition on high-level plans via explicit attention, hierarchical gating, or multi-head mechanisms. For example, hierarchical attention combines word-level and sentence-level attention over outlined content, computing context vectors for each decoding step (Drissi et al., 2018).
• Self-attention is extended with scale-specific heads and gating to capture both local and long-range narrative dependencies, as in the gated multi-scale attention used for story realization (Fan et al., 2018).

Hybridization strategies, such as fusion models that combine pretrained and learned networks, further increase prompt relevance and content control (Fan et al., 2018). Recent architectures incorporate memory modules (e.g., temporal knowledge graphs) that track story entities and actions over time to inform context and prevent contradictions (Wang et al., 18 Dec 2024, Li et al., 3 Jun 2025).

3. Methods of High-level Abstraction and Planning

The "outline" role is instantiated as:

Abstraction Type	Description	Example Systems
Sentence Outlines	Sequence of major plot sentences or "beats"	(Drissi et al., 2018, Wang et al., 2020)
Anchor Words (Latent Plans)	One anchor or topic word per sentence, often as a latent variable	(Jhamtani et al., 2020)
Predicate–Argument Structures	SRL-based sequence of verbs and argument roles (PAS)	(Fan et al., 2019)
Event Graphs/Chains	Corpus-derived chains of causally related events	(Chen et al., 2021)
SVO Plot Nodes	Linguistically-grounded subject–verb–object structures	(Li et al., 3 Jun 2025)
Dynamic Hierarchical Outlines	Multi-level outline fusing theory-driven writing stages	(Wang et al., 18 Dec 2024)

These abstractions are typically constructed either via preprocessing (e.g., SRL, TextRank, RAKE) or induced in a latent/unsupervised fashion, as in the case of anchor word selection (Jhamtani et al., 2020, Fan et al., 2019).

4. Training Procedures, Losses, and Consistency Enhancements

Hierarchical models are trained with stacked or joint objectives:

• Stage-wise Training: Each generation stage (outline, PAS, entity replacement, etc.) is trained separately, usually minimizing cross-entropy (Drissi et al., 2018, Wang et al., 2020, Fan et al., 2019).
• Joint/End-to-End Training: Some models train stages jointly, for instance by maximizing a combined ELBO or policy-gradient-based objective (Jhamtani et al., 2020, Huang et al., 2018).
• Auxiliary Losses: Specific tasks, such as coreference consistency and discourse relation modeling, are incorporated to encourage long-range entity tracking and discourse coherence. Coreference supervision encourages decoder self-attention to focus on antecedents with matching entity labels, and discourse relation modeling provides explicit signals for transitions between sentences (Wang et al., 2020).
• Reinforcement Learning: Applied to optimize non-differentiable metrics and story-level structure, often with rewards computed from story-level CIDEr or planning-specific criteria (Huang et al., 2018).

Ablations consistently show that including hierarchy and these auxiliary losses significantly improves event diversity, entity consistency, and temporal coherence, while reducing repetition and factual inconsistency (Fan et al., 2019, Wang et al., 2020).

5. Evaluation Protocols and Empirical Findings

Both automatic and human evaluations are used:

• Perplexity: Hierarchical models often show perplexity reduction over flat LMs; e.g., outline→article with hierarchical attention achieves 20.5 vs. 31.0 for flat, prompt-conditioned baselines (Drissi et al., 2018). However, perplexity does not reliably track global coherence as perceived by humans.
• Human Ratings: Judges score stories on dimensions such as logical coherence, relevance, and grammaticality. Hierarchical models are generally preferred: in direct comparisons, up to a 2:1 preference ratio has been observed in favor of hierarchical generation over non-hierarchical baselines (Fan et al., 2018, Wang et al., 2020).
• Diversity and Consistency Metrics: Metrics such as Distinct-n, anchor control rates, number of coreference chains, and event diversity are used to quantify the richness and consistency of generated narratives (Jhamtani et al., 2020, Wang et al., 2020).
• Conflict Rate and Entity Consistency: Recent models use explicit rates of temporal conflicts (contradictory events on the story timeline) and entity-specific metrics to measure long-form consistency gains achieved by memory-enhanced outlines (Wang et al., 18 Dec 2024).
• SCT and Planning Quality: Planning algorithms are also evaluated by story cloze test accuracy and human annotation of logicality and causal structure (Chen et al., 2021).

A consistent observation is that improvements in sequence-level likelihood and token-level perplexity may not correspond to gains in narrative coherence as judged by humans (Drissi et al., 2018).

6. Advances in Dynamic Planning, Memory, and Structural Integration

Contemporary systems extend conventional hierarchy in several directions:

• Dynamic Outlining and Adaptive Planning: Rather than generating a fixed outline prior to text, models such as DOME update detailed outlines dynamically during generation, conditioned on generated content and past narrative (retrieval from temporal knowledge graphs) (Wang et al., 18 Dec 2024).
• Memory Modules: Temporal knowledge graphs, narrative entity knowledge graphs (NEKG), and memory-augmentation are integrated to track character states, actions, and facts, reducing contradiction and increasing narrative continuity (Wang et al., 18 Dec 2024, Li et al., 3 Jun 2025).
• Structural Filters and Review Mechanisms: The use of NEKG in STORYTELLER permits dynamic review and editing of proposed plot events, enforcing event non-redundancy and continuity during plot planning (Li et al., 3 Jun 2025).
• Multi-Granularity and Coarse-to-Fine Realization: Multiple intermediate stages, such as PAS, anonymized text, and entity realization, allow for modular and interpretable control over event diversity and character tracking (Fan et al., 2019).

Empirical results show that such mechanisms robustly increase long-n-gram diversity, reduce conflicts, and produce higher-rated stories by both automatic and human metrics (Wang et al., 18 Dec 2024, Li et al., 3 Jun 2025).

7. Implications, Limitations, and Future Directions

Hierarchical neural story generation architectures directly address core issues of narrative logic, topical focus, and long-range consistency that afflict flat, left-to-right LMs. By dividing the generation process into structured sub-tasks—often modeled on human writing practices—they provide explicit control over plot structure, event progression, and character evolution.

Current limitations include sensitivity to the quality of high-level plans (outlines, event chains), exposure bias between planning and realization stages, and the need for specialized preprocessing (SRL, event extraction, topic modeling). Models also reveal a gap between automatic metrics (e.g., perplexity) and perceived narrative quality, emphasizing the necessity of human-centered evaluation protocols.

Recent work suggests that fusing planning, memory, and dynamic feedback mechanisms provides new avenues for robust, interpretable, and controllable narrative generation, with demonstrated gains in both fluency and coherence for long-form storytelling (Wang et al., 18 Dec 2024, Li et al., 3 Jun 2025). Future research directions include unsupervised or latent hierarchical induction, tighter planner-realizer integration, coverage-aware decoding, reinforcement objectives for coherence, and modular plug-ins for third-order narrative phenomena such as theme, style, or genre control.