Story Generator Structure

• Story generator structure is a framework that decomposes narrative creation into modular, hierarchical components controlling planning, textual realization, and stylistic cues.
• It utilizes methods such as two-stage pipelines, semantic abstractions, and multi-agent feedback to enhance global planning and coherence in generated stories.
• This approach integrates external knowledge, memory modules, and interaction loops to produce narratives with consistent entities, diverse plots, and thematic fidelity.

A story generator structure refers to the deliberate decomposition of the story generation process into multiple explicit modules or abstraction layers, each controlling a facet of narrative logic, global coherence, or stylistic realization. Contemporary research demonstrates that such explicit decompositions are critical for generating stories that are coherent at scale, maintain entity and plot consistency, and achieve sufficient thematic and stylistic diversity.

1. Modular and Hierarchical Architectures

Recent story generation systems consistently employ hierarchical or modular pipelines. Canonical pipelines include two-stage architectures (outline→story), coarse-to-fine decompositions (semantic abstraction→surface realization), and multi-agent or collaborative frameworks. Modular separation allows for explicit global planning followed by local realization, or interactive feedback loops between planning and revision.

For example, "Plan-And-Write" implements a clear two-stage pipeline—a Planner produces a sequence of storyline keywords or events, and a Writer conditions on this plan to realize coherent text (Yao et al., 2018). Similar decompositions appear in content-planning pipelines that factorize $p(y \mid x)$ as $p(z \mid x)\cdot p(y \mid z, x)$, where $z$ is a structured plot or outline (Goldfarb-Tarrant et al., 2020, Wang et al., 2020).

Pipeline variants:

Architecture	Planning Layer	Realization Layer
Plan-and-write	Keyword or event sequence	Seq2Seq conditional LM w/ attention
Consistency-enhanced	Abstract "outline" (sentence/keyword)	Transformer-Decoder w/ outline context
Multi-agent frameworks	Distributed agent event planning	Multi-agent writing/feedback
Predicate-argument	SRL frames + placeholders (SRL/NER/Coref)	Surface realizer & entity refiller

This modularization contrasts with pure left-to-right LM generation, affording structural control and tractable intermediate objectives (Xia et al., 19 Jun 2025, Fan et al., 2019).

2. Planning Representations and Strategies

Planning modules employ diverse representations: sequences of keywords, predicate–argument structures, events annotated with time/role/object, subject–verb–object (SVO) triples, or even logic-based and ASP-encoded narrative functions.

• In static planning, the full story structure (e.g., a keyword sequence or graph) is generated before any text realization (Yao et al., 2018, Wang et al., 1 Jun 2024).
• In dynamic/interleaved planning, planning and realization alternate at each step, allowing finer coupling and immediate feedback for each action (Yao et al., 2018).
• Predicate-argument or SRL-based plans encode story events as a sequence of verb frames or SRL tuples (Fan et al., 2019), allowing explicit control over event diversity and arguments.
• SVO triplets structurally enforce event atomicity and provide cross-event entity linking, facilitating consistent plot node expansion (Li et al., 3 Jun 2025).

Formally, plans are generated autoregressively: $p(z|x)=\prod_{i} p(z_i|z_{<i},x)$ for keyword, sentence, or event sequence plans, with further augmentation by rescoring models or knowledge-graph constraints (Goldfarb-Tarrant et al., 2020, Wang et al., 2022, Shi et al., 5 Aug 2025).

3. Realization and Surface Generation

Generation modules typically employ neural sequence-to-sequence models (LSTM, GRU, Transformer, ConvS2S), with explicit attention mechanisms over the plan or outline. In hierarchical generators, the plan is encoded (via BiLSTM, Transformer, graph neural net), and story realization proceeds by conditioning on the encoded plan as context.

For example: $p(y|z, x) = \prod_{t=1}^n p(y_t| y_{<t}, z, x)$ Realization may include attention over multiple conditioning sources, copy mechanisms (e.g., pointer-generator, entity refiller), or fusion of multiple model outputs (Fan et al., 2019, Fan et al., 2018).

Advanced systems augment realization with:

• Coreference loss: guides attention weights over prior mentions to improve pronoun consistency (Wang et al., 2020).
• Discourse modeling loss: auxiliary connective or discourse relation classification to enforce local coherence (Wang et al., 2020).
• Graph-based attention: integrates structured knowledge or plot graphs into encoder states (Wang et al., 2022, Li et al., 3 Jun 2025).

4. Integrating Knowledge, Memory, and Interaction

Knowledge-enhanced generators inject external knowledge sources at the planning or realization stage. Structured knowledge can be encoded as:

• Knowledge graphs (concepts, events): used to ground plans and inform event selection (Wang et al., 2022, Shi et al., 5 Aug 2025).
• Memory modules: maintain both long-term theme representations and short-term outline histories to avoid theme drift, with retrieval via embedding similarity and top-K selection (Shi et al., 5 Aug 2025).
• Multi-agent and critic–writer feedback: collaborative or adversarial modules (e.g., writer–reader simulators, event validators) that revise or filter drafts to ensure logic and closure (Xia et al., 19 Jun 2025, Shi et al., 5 Aug 2025, Chen et al., 13 Oct 2025).

Procedural and game-based generators further align story structure to emotional arcs, mapping event difficulty and content as a function of global narrative valence, validated with sentiment classifiers (Wen et al., 4 Aug 2025).

5. Evaluation Metrics and Empirical Outcomes

Objective evaluation includes:

• Inter- and intra-story repetition rate (trigram overlap) for diversity (Yao et al., 2018, Goldfarb-Tarrant et al., 2020).
• BLEU and Distinct-n for content quality and lexical diversity (Yao et al., 2018, Goldfarb-Tarrant et al., 2020, Jin et al., 2022).
• Automatic and human judgments for coherence, on-topic fidelity, interestingness, and overall quality (Yao et al., 2018, Wang et al., 2020, Xia et al., 19 Jun 2025, Shi et al., 5 Aug 2025).
• Coreference and discourse relation consistency markers (Wang et al., 2020).
• Pairwise human preferences, win rates, Brier score, and Cohen's Kappa for annotation agreement (Li et al., 3 Jun 2025, Xia et al., 19 Jun 2025).
• Novel story-specific metrics, such as length-targeted formulas and plot-structure token ratios (Xia et al., 19 Jun 2025, Goldfarb-Tarrant et al., 2020, Fan et al., 2018).

Empirically, systems with explicit hierarchical planning, dynamic knowledge graphs, and feedback mechanisms outperform single-stage LMs along all major axes: coherence, creativity, entity/event diversity, and thematic tightness (Yao et al., 2018, Fan et al., 2019, Xia et al., 19 Jun 2025, Shi et al., 5 Aug 2025).

6. Implementation Schemas and Best Practices

Implementation details are rigorously described for each architecture:

• Neural architectures: hybrid BiLSTM/GRU (planning), Transformer decoder-only (outline/expander), pointer-generator with coverage (ending generator), convolutional seq2seq with gated multi-scale attention (hierarchical generator) (Yao et al., 2018, Wang et al., 2020, Zhao et al., 2019, Fan et al., 2018).
• Auxiliary feature extraction: storylines via unsupervised RAKE (Yao et al., 2018) or abstract extraction; event graphs and SRL via OpenIE/SRL pipelines (Fan et al., 2019).
• Training regimes: pipeline or joint (e.g., outline-then-story), cross-entropy loss with auxiliary terms for coreference/discourse (Wang et al., 2020, Yao et al., 2018).
• Decoding: greedy, beam search, sampling, with beam sizes (5–20) tailored to plan and realization stages (Yao et al., 2018, Wang et al., 2020).
• Hyperparameters: embedding dimensions (100–500 for keywords, up to 1000 for hidden), optimizers (SGD, Adam), dropout ranges (Yao et al., 2018, Wang et al., 2020).

Block-diagram representations, pseudocode, and explicit update rules (e.g., memory and interaction feedback) are provided to guarantee reproducibility (Yao et al., 2018, Shi et al., 5 Aug 2025, Li et al., 3 Jun 2025).

7. Contemporary Trends and Extensions

Modern generators increasingly leverage:

• Multi-agent simulation for emergent, bottom-up event generation, using agent LLMs and environment state machines (Chen et al., 13 Oct 2025).
• Structured frameworks for user control (e.g., TaleFrame's E/V/R/O units edited via HCI) and fine-grained JSON-to-story pipelines (Wang et al., 2 Dec 2025).
• Hybrid neurosymbolic architectures (ASP + LLM) for outline diversity and adherence to symbolic narrative constraints (Wang et al., 1 Jun 2024).
• RL-based learning of reasoning, where next-chapter generation is enhanced by plan tokens validated by likelihood improvement (Gurung et al., 28 Mar 2025).
• Game and multimodal narrative generation, with structural alignment to emotional arcs and universal story templates (Wen et al., 4 Aug 2025, Chen et al., 2023).

Limitations discussed include evaluation–judgment correlation gaps, the brittleness of rigid planning, and the need for more adaptive, interactive, and knowledge-rich modules. Future directions emphasize interactive editing, blending top-down and bottom-up structure, and integrating richer memory, commonsense, and discourse models.

---

References:

• "Plan-And-Write: Towards Better Automatic Storytelling" (Yao et al., 2018)
• "Consistency and Coherency Enhanced Story Generation" (Wang et al., 2020)
• "Long Story Generation via Knowledge Graph and Literary Theory" (Shi et al., 5 Aug 2025)
• "Content Planning for Neural Story Generation with Aristotelian Rescoring" (Goldfarb-Tarrant et al., 2020)
• "Guiding and Diversifying LLM-Based Story Generation via Answer Set Programming" (Wang et al., 1 Jun 2024)
• "StoryWriter: A Multi-Agent Framework for Long Story Generation" (Xia et al., 19 Jun 2025)
• "Strategies for Structuring Story Generation" (Fan et al., 2019)
• "Automated Story Generation as Question-Answering" (Castricato et al., 2021)
• "Open-world Story Generation with Structured Knowledge Enhancement: A Comprehensive Survey" (Wang et al., 2022)
• "Hierarchical Neural Story Generation" (Fan et al., 2018)
• "A Customizable Generator for Comic-Style Visual Narrative" (Chen et al., 2023)
• "All Stories Are One Story: Emotional Arc Guided Procedural Game Level Generation" (Wen et al., 4 Aug 2025)
• "Learning to Reason for Long-Form Story Generation" (Gurung et al., 28 Mar 2025)
• "From Plots to Endings: A Reinforced Pointer Generator for Story Ending Generation" (Zhao et al., 2019)
• "STORYTELLER: An Enhanced Plot-Planning Framework for Coherent and Cohesive Story Generation" (Li et al., 3 Jun 2025)
• "StoryBox: Collaborative Multi-Agent Simulation for Hybrid Bottom-Up Long-Form Story Generation Using LLMs" (Chen et al., 13 Oct 2025)
• "Generating Different Story Tellings from Semantic Representations of Narrative" (Rishes et al., 2017)
• "Plot Writing From Pre-Trained LLMs" (Jin et al., 2022)
• "TaleFrame: An Interactive Story Generation System with Fine-Grained Control and LLMs" (Wang et al., 2 Dec 2025)