Strategies for Structuring Story Generation

Angela Fan, Mike Lewis, Yann N. Dauphin

Abstract

Writers generally rely on plans or sketches to write long stories, but most current language models 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.

Introduction

While modern language models are excellent at maintaining local grammar, they often fail at "global coherence"—the ability to maintain a consistent plot or character identity over several hundred words. Standard models generate text from left to right, word by word, making it difficult to capture high-level interactions between plot points. This research introduces a "coarse-to-fine" framework that mimics the human writing process by sketching an outline before drafting the final text.

Methodology: Coarse-to-Fine Decomposition

The proposed framework decomposes story generation into three distinct stages to manage long-range dependencies and entity consistency:

  1. Action Sequence Generation: Instead of raw text, the model generates a sequence of "predicates and arguments" using Semantic Role Labeling (SRL). To prevent the model from repeating generic actions, a specialized Verb Attention Mechanism is introduced in the transformer decoder to focus on previously generated verbs.
  2. Entity Anonymization: The model generates a "surface realization" where characters and locations are replaced by placeholders like ent0 and ent1. This allows the model to focus on narrative flow without the difficulty of generating rare names.
  3. Entity Reference Generation: A specialized sub-word model replaces placeholders with actual names, pronouns, or descriptions using a pointer-copy mechanism to ensure consistent reuse of names.
The proposed coarse-to-fine generation pipeline, moving from a prompt to a structured action plan, then to an anonymized draft, and finally the full story.

Experimental Results

The models were evaluated on the WRITINGPROMPTS dataset, which contains 300,000 challenging long-form stories from Reddit.

Human Preference

In blind tests, human judges preferred stories from the SRL + Coreference model over the state-of-the-art "Fusion" baseline approximately 59.7% of the time. The structured approach consistently outperformed other hierarchical methods such as keyword extraction or summarization.

Human preference results in blind tests, showing the proposed SRL + Coref model outperforming various baselines including Fusion and SRL + NER.

Narrative Diversity and Coherence

  • Verb Diversity: The proposed model increased the number of unique verbs per story from 10.3 (baseline) to 15.9, moving closer to the human average of 34.0.
  • Entity Consistency: Coreference-based anonymization led to more sophisticated entity chains, successfully mixing names and pronouns rather than repeating a single name.
  • Tractability: The action plan generation achieved a significantly lower negative log-likelihood (2.72) compared to keyword extraction (6.92), suggesting SRL is a more tractable representation for story structure.

Conclusion

By abstracting over actions and entities, this coarse-to-fine framework significantly improves the coherence and diversity of generated narratives. While the model is better at consistency, future work is needed to incorporate explicit world knowledge and common-sense reasoning to ensure that action plans are not only diverse but also logically sound.

Summary updated

Abstract

Writers generally rely on plans or sketches to write long stories, but most current language models 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.
BibTeX
@inproceedings{fan2019strategies,
  author = {Angela Fan and Mike Lewis and Yann N. Dauphin},
  title = {{Strategies for Structuring Story Generation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
  url = {https://openreview.net/forum?id=HY1zxLZ7_BM},
}
Figure