Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
110 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
44 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

Strategies for Structuring Story Generation (1902.01109v2)

Published 4 Feb 2019 in cs.CL

Abstract: Writers generally rely on plans or sketches to write long stories, but most current LLMs generate word by word from left to right. We explore coarse-to-fine models for creating narrative texts of several hundred words, and introduce new models which decompose stories by abstracting over actions and entities. The model first generates the predicate-argument structure of the text, where different mentions of the same entity are marked with placeholder tokens. It then generates a surface realization of the predicate-argument structure, and finally replaces the entity placeholders with context-sensitive names and references. Human judges prefer the stories from our models to a wide range of previous approaches to hierarchical text generation. Extensive analysis shows that our methods can help improve the diversity and coherence of events and entities in generated stories.

User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (3)
  1. Angela Fan (49 papers)
  2. Mike Lewis (78 papers)
  3. Yann Dauphin (24 papers)
Citations (208)

Summary

An Analysis of Strategies for Structuring Story Generation

The paper "Strategies for Structuring Story Generation" by Fan, Lewis, and Dauphin explores an innovative approach to narrative text generation by examining decomposition strategies that can enhance the structure and coherence of generated stories. Traditional LLMs typically generate text sequentially at the word level, a method that can lead to disjointed narratives and inconsistent characters across a story's length. This work investigates a coarse-to-fine generation methodology that breaks down the story creation process into manageable subproblems to improve overall narrative quality.

Central to the authors' approach is the decomposition of a story into a sequence of predicate-argument structures that emphasize the logical flow of actions within a narrative. This abstraction facilitates easier modeling of dependencies across events, allowing the model to generate sequences of predicates and arguments before fleshing out these sequences with placeholders for entities. The placeholders are later replaced with context-appropriate names and references, ensuring entities are coherent with the overall story context. This structured generation process draws inspiration from classical text generation models, which were based on a multilevel view of text structure.

The empirical evaluation is thorough, replete with both human judgments and automated metrics. The model was trained on a substantial corpus of 300k stories from the WritingPrompts dataset, which provided a diverse range of narratives and allowed for robust assessment of the proposed method's efficacy. The method demonstrates a capability to enhance diversity in narrative action and improve the coherence of entities throughout a story, which are significant challenges in story generation tasks. Human evaluators consistently favored stories generated with this novel decomposition approach over alternatives, indicating the authors' strategy yields a qualitative improvement in narrative engagement.

The implementation of specialized approaches at different stages—particularly in the modeling of actions and entities—leads to improved outcomes. By utilizing concepts such as Semantic Role Labeling (SRL) and coreference resolution, the model is adept at maintaining the logical progression and coherence of stories, a key advancement over straightforward left-to-right generation models. Action sequences are enhanced via verb-attention mechanisms, resulting in a broader variety of action verbs and reduced repetition.

Additionally, entity modeling is refined through the introduction of placeholder tokens, which address the typical neural LLM problem of generating rare or novel entity names. The use of character-level modeling and pointer mechanisms supplement this strategy, enabling the generation of contextually relevant and diverse entity names that align closely with human creativity observed in story writing.

While the results are promising, the paper also gestures towards potential directions for further research. The decomposition approach reveals the complexity and the distinct challenges inherent in each stage of story generation, thus inviting continued exploration into how different layers of predication and referencing can be fine-tuned or expanded upon.

In synthesis, the paper adds a valuable contribution to the field of natural language processing, particularly in narrative text generation. Its method enhances story coherence and variety, addressing key weaknesses in traditional LLMs. Future developments could see this approach applied more broadly, potentially improving narrative quality across diverse applications such as automated storytelling, scriptwriting, and educational content generation. The insights drawn from this research could inform new models that are capable of generating stories that more closely align with human cognitive expectations of coherence, thematic variety, and character consistency.