Generating Long-form Story Using Dynamic Hierarchical Outlining with Memory-Enhancement

TL;DR

This work tackles the challenge of cohering long-form story generation by integrating a Dynamic Hierarchical Outline (DHO) with a Memory-Enhancement Module (MEM) built on Temporal Knowledge Graphs, plus a Temporal Conflict Analyzer to automatically assess coherence. The key contributions are the DHO that fuses planning and writing under novel-writing theory with dynamic outline expansion, the MEM that stores and retrieves concise, query-relevant content to reduce conflicts, and an automatic temporal conflict metric for evaluation. Empirical results show improvements in both macro-level plot coherence ($C^{Plot}$) and context coherence ($C^{Context}$) over state-of-the-art baselines, with demonstrated scalability across large and smaller LLMs and manageable computational overhead. The approach holds practical significance for automatic long-form storytelling in novel writing and interactive systems, while noting limitations such as reliance on prompts and the cost of large-scale LLM usage.

Abstract

Long-form story generation task aims to produce coherent and sufficiently lengthy text, essential for applications such as novel writingand interactive storytelling. However, existing methods, including LLMs, rely on rigid outlines or lack macro-level planning, making it difficult to achieve both contextual consistency and coherent plot development in long-form story generation. To address this issues, we propose Dynamic Hierarchical Outlining with Memory-Enhancement long-form story generation method, named DOME, to generate the long-form story with coherent content and plot. Specifically, the Dynamic Hierarchical Outline(DHO) mechanism incorporates the novel writing theory into outline planning and fuses the plan and writing stages together, improving the coherence of the plot by ensuring the plot completeness and adapting to the uncertainty during story generation. A Memory-Enhancement Module (MEM) based on temporal knowledge graphs is introduced to store and access the generated content, reducing contextual conflicts and improving story coherence. Finally, we propose a Temporal Conflict Analyzer leveraging temporal knowledge graphs to automatically evaluate the contextual consistency of long-form story. Experiments demonstrate that DOME significantly improves the fluency, coherence, and overall quality of generated long stories compared to state-of-the-art methods.

Figures (30)

• Figure 1: Illustration and comparison of three strategies of long-form story generation. (a) applying a fixed hierarchical outline to guide the story generation, which is challenging to adapt to the uncertainty in story creation. (b) adapting to the uncertainty through interaction with humans, which allows for flexibility but lacks a high-level storyline to guide story development. Our method shown in (c) aims to enhance story coherence from both plot and expression and enjoys the advantages of the former two strategies to improve the coherence of the plot and reduce contextual conflicts.
• Figure 2: General diagram of proposed DOME. The story generation process is divided into several stages based on the amount of rough outline. At stage $i$, we expand rough outline $i$ into several detailed outlines based on the relevant content provided by MEM, and these outlines generate partial story sequentially based on their relevant content querying from MEM. Every generated partial story is stored in a temporal Knowledge graph for the following query.
• Figure 3: Five stages novel writing theory from Joseph Campbell TheoryFive.
• Figure 4: The details of querying for the relevant content. Notably the last arrow points to the historical information semantically related to the input content.
• Figure 5: The criterion of filtering on semantic relevance.