Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Returning to the Start: Generating Narratives with Related Endpoints

Anneliese Brei, Chao Zhao, Snigdha Chaturvedi

TL;DR

The paper addresses the challenge of generating narratives with strong narrative closure by enforcing related endpoints and infilling the middle of the story. It introduces RENarGen, a framework that handles endpoints and infilling with two pathways: (i) LM-based generation using semantic relatedness to relate the start to a stop, and (ii) LLM-based generation using diverse prompting strategies that encode various notions of relatedness, including erotetic closure and entailment. Through a two-component design—Endpoint Generator and Story Infiller—RENarGen demonstrates improved endpoint relatedness and overall coherence over baselines, validated by both automatic metrics and human judgments on the ROCStories dataset. The work shows that narratology-inspired bookending can meaningfully improve automated narrative generation, with interactive capabilities for LM users and flexible prompting for LLMs, indicating practical impact for controllable narrative synthesis. Overall, RENarGen advances the state of automatic story generation by formalizing related-endpoint bookending and validating its benefits across multiple evaluation dimensions.

Abstract

Human writers often bookend their writing with ending sentences that relate back to the beginning sentences in order to compose a satisfying narrative that "closes the loop." Motivated by this observation, we propose RENarGen, a controllable story-generation paradigm that generates narratives by ensuring the first and last sentences are related and then infilling the middle sentences. Our contributions include an initial exploration of how various methods of bookending from Narratology affect language modeling for stories. Automatic and human evaluations indicate RENarGen produces better stories with more narrative closure than current autoregressive models.

Returning to the Start: Generating Narratives with Related Endpoints

TL;DR

The paper addresses the challenge of generating narratives with strong narrative closure by enforcing related endpoints and infilling the middle of the story. It introduces RENarGen, a framework that handles endpoints and infilling with two pathways: (i) LM-based generation using semantic relatedness to relate the start to a stop, and (ii) LLM-based generation using diverse prompting strategies that encode various notions of relatedness, including erotetic closure and entailment. Through a two-component design—Endpoint Generator and Story Infiller—RENarGen demonstrates improved endpoint relatedness and overall coherence over baselines, validated by both automatic metrics and human judgments on the ROCStories dataset. The work shows that narratology-inspired bookending can meaningfully improve automated narrative generation, with interactive capabilities for LM users and flexible prompting for LLMs, indicating practical impact for controllable narrative synthesis. Overall, RENarGen advances the state of automatic story generation by formalizing related-endpoint bookending and validating its benefits across multiple evaluation dimensions.

Abstract

Human writers often bookend their writing with ending sentences that relate back to the beginning sentences in order to compose a satisfying narrative that "closes the loop." Motivated by this observation, we propose RENarGen, a controllable story-generation paradigm that generates narratives by ensuring the first and last sentences are related and then infilling the middle sentences. Our contributions include an initial exploration of how various methods of bookending from Narratology affect language modeling for stories. Automatic and human evaluations indicate RENarGen produces better stories with more narrative closure than current autoregressive models.
Paper Structure (31 sections, 2 figures, 2 tables)

This paper contains 31 sections, 2 figures, 2 tables.

Table of Contents

  1. Introduction
  2. RENarGen
  3. Endpoint Generator for LMs
  4. Story Infiller for LMs
  5. RENarGen for LLMs
  6. Empirical Evaluation
  7. Dataset
  8. Automatic Evaluation
  9. Human Evaluation of Quality
  10. Human Evaluation of Interactivity
  11. Conclusion
  12. Acknowledgement
  13. Limitations
  14. Ethics Statement
  15. Example Generation with RENarGen
  16. ...and 16 more sections

Figures (2)

  • Figure 1: Stories with related start and stop sentences (Story 1, generated by RENarGen) provide better narrative closure than stories with unrelated endpoints (Story 2, generated by GPT-2 baseline).
  • Figure 2: Proposed RENarGen framework. Box 1: Scheme for LMs. Given input start, the Phrase Generator produces a phrase list of relatable words; using this list, the Stop Generator outputs the stop. The Story Infiller infills middle sentences by iteratively determining the next best location for a new sentence and generating a sentence. A sample step-by-step story generation is given in Appendix \ref{['sec:gen_example']}. Box 2: Scheme for LLMs. Given input start, Endpoint Generater chooses one of six methods to generate the stop. The Story infiller uses the start and stop to generate all infills. After data cleaning, all components are concatenated into the final full story.