Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Affective and Dynamic Beam Search for Story Generation

Tenghao Huang, Ehsan Qasemi, Bangzheng Li, He Wang, Faeze Brahman, Muhao Chen, Snigdha Chaturvedi

TL;DR

This work tackles automatic generation of engaging stories by introducing AffGen, a decoding-time framework that combines Dynamic Beam Sizing and Affective Reranking to inject intriguing twists while maintaining narrative coherence. Twist placement is data-driven, sampling from a distribution of where a climax or turning point is likely, and the twist is generated using a neural LM with a contextual bandit to select beam sizes and an affective reranker to optimize arousal and emotional dynamics. Empirical results on ROCStories show AffGen achieves higher arousal and perceived interestingness with coherent text, supported by human judgments that favor AffGen over baselines like GPT-3 and ChatGPT in key affective and engagement measures. Ablation, longer-narrative expansion, and qualitative analyses further demonstrate the value of dynamic exploration and affective re-scoring for computational creativity, while acknowledging limitations such as potential repetition and reliance on single-twist designs.

Abstract

Storytelling's captivating potential makes it a fascinating research area, with implications for entertainment, education, therapy, and cognitive studies. In this paper, we propose Affective Story Generator (AffGen) for generating interesting narratives. AffGen introduces "intriguing twists" in narratives by employing two novel techniques-Dynamic Beam Sizing and Affective Reranking. Dynamic Beam Sizing encourages less predictable, more captivating word choices using a contextual multi-arm bandit model. Affective Reranking prioritizes sentence candidates based on affect intensity. Our empirical evaluations, both automatic and human, demonstrate AffGen's superior performance over existing baselines in generating affectively charged and interesting narratives. Our ablation study and analysis provide insights into the strengths and weaknesses of AffGen.

Affective and Dynamic Beam Search for Story Generation

TL;DR

This work tackles automatic generation of engaging stories by introducing AffGen, a decoding-time framework that combines Dynamic Beam Sizing and Affective Reranking to inject intriguing twists while maintaining narrative coherence. Twist placement is data-driven, sampling from a distribution of where a climax or turning point is likely, and the twist is generated using a neural LM with a contextual bandit to select beam sizes and an affective reranker to optimize arousal and emotional dynamics. Empirical results on ROCStories show AffGen achieves higher arousal and perceived interestingness with coherent text, supported by human judgments that favor AffGen over baselines like GPT-3 and ChatGPT in key affective and engagement measures. Ablation, longer-narrative expansion, and qualitative analyses further demonstrate the value of dynamic exploration and affective re-scoring for computational creativity, while acknowledging limitations such as potential repetition and reliance on single-twist designs.

Abstract

Storytelling's captivating potential makes it a fascinating research area, with implications for entertainment, education, therapy, and cognitive studies. In this paper, we propose Affective Story Generator (AffGen) for generating interesting narratives. AffGen introduces "intriguing twists" in narratives by employing two novel techniques-Dynamic Beam Sizing and Affective Reranking. Dynamic Beam Sizing encourages less predictable, more captivating word choices using a contextual multi-arm bandit model. Affective Reranking prioritizes sentence candidates based on affect intensity. Our empirical evaluations, both automatic and human, demonstrate AffGen's superior performance over existing baselines in generating affectively charged and interesting narratives. Our ablation study and analysis provide insights into the strengths and weaknesses of AffGen.
Paper Structure (22 sections, 8 equations, 4 figures, 10 tables)

This paper contains 22 sections, 8 equations, 4 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Problem statement
  4. Controlled Affective Story Generator
  5. Position of the intriguing twist
  6. Base Storyteller
  7. Generating Intriguing Twist
  8. Empirical Evaluation
  9. Experimental Setup
  10. Automatic Evaluation
  11. Human Evaluation
  12. Comparison with ChatGPT
  13. Ablation Study
  14. Expansion to Longer Narratives
  15. Dynamic Beam Sizing
  16. ...and 7 more sections

Figures (4)

  • Figure 1: Two example stories. Story 1 is an interesting story with an intriguing twist (highlighted in orange color) that was produced by AffGen using dynamic beam sizing. Story 2 is a relatively less interesting story with a straightforward and predictable plot.
  • Figure 2: Average beam size used to generate at different positions of a sentence.
  • Figure 3: Transitional probability between beam sizes.
  • Figure 4: Screenshot of the AMT survey used for human evaluations in \ref{['sec:human_eval']}