Strategies for Structuring Story Generation

Angela Fan, Mike Lewis, Yann N. Dauphin

Abstract

Writers often 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

Standard language models (LMs) typically generate text one word at a time, from left to right. While effective for short-form content, this approach often fails in long-form storytelling. Two primary issues emerge: a lack of logical plot progression and poor character consistency. LMs frequently lose the "thread" of a narrative, leading to repetitive or contradictory events. Furthermore, character names—often "rare tokens" in a vocabulary—are difficult for LMs to maintain consistently, frequently resulting in incorrect pronoun usage or name switching.

This research argues that human storytelling is inherently a planned process. By integrating a planning phase into neural architectures, we can move beyond simple word-by-word prediction toward more coherent, structured narratives.

Methodology

The proposed framework utilizes a three-stage "coarse-to-fine" decomposition inspired by classical Natural Language Generation (NLG) pipelines.

The three-stage hierarchical generation process: action planning, anonymized narrative generation, and entity realization.

1. Action Plan Generation

Instead of generating raw text immediately, the model first produces a sequence of predicates and arguments using Semantic Role Labeling (SRL). For instance, a sentence like "John ate the apple" is abstracted to [ate] <ARG0> ent0 <ARG1> apple. To ensure logical action sequences, a Verb Attention Mechanism was introduced, where a specific head in the transformer-style decoder attends exclusively to previously generated verbs.

2. Entity Anonymized Story

The model then generates the full narrative flow but replaces all character names and entities with placeholders (e.g., ent0, ent1). This abstraction allows the model to focus on narrative logic and prose without the distraction of managing specific entity names.

3. Entity Reference Generation

A final specialized model replaces placeholders with context-appropriate names or pronouns. This stage employs a character-level model and a pointer mechanism. The character-level approach facilitates the creation of unique names, while the pointer mechanism ensures that if ent0 is identified as "Bilbo," the model consistently uses that name or appropriate pronouns later in the text.

Dataset

The models were trained on the WRITINGPROMPTS dataset, comprising approximately 300,000 story premises and long-form stories (averaging 734 words) from Reddit. This dataset provides a significant challenge due to the length and creative diversity of the entries.

Results and Analysis

The study revealed several key insights regarding hierarchical generation:

  • Abstraction Efficiency: The "Action Plan" (SRL) demonstrates a lower negative log-likelihood than raw text, suggesting it is mathematically easier for models to learn the logic of actions than complex surface-level prose.
  • Verb Diversity: SRL-based models produced a significantly wider variety of unique verbs compared to standard LMs, which often rely on generic actions.
  • Human Preference: In blind A/B testing, human judges significantly preferred the structured models.
Human preference results showing the SRL + Coref model outperforming previous state-of-the-art approaches.

As shown in the evaluation, the SRL + Coreference model was preferred over the previous state-of-the-art "Fusion" model nearly 60% of the time. The model also showed marked improvements in "coreference chains," maintaining character consistency over hundreds of words.

Conclusion

While this coarse-to-fine approach represents a significant advancement in narrative coherence, a "human gap" remains in terms of action and entity diversity. Current limitations include a reliance on automated SRL tools, which can propagate errors, and a lack of global plotting for thematic arcs or specific endings. Future research will explore higher-level abstractions to plan emotional arcs and climactic structures before generating text.

Summary updated

Abstract

Writers often 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