Modeling Protagonist Emotions for Emotion-Aware Storytelling

Faeze Brahman, Snigdha Chaturvedi

Abstract

Emotions and their evolution play a central role in creating a captivating story. In this paper, we present the first study on modeling the emotional trajectory of the protagonist in neural storytelling. We design methods that generate stories that adhere to given story titles and desired emotion arcs for the protagonist. Our models include Emotion Supervision (EmoSup) and two Emotion-Reinforced (EmoRL) models. The EmoRL models use special rewards designed to regularize the story generation process through reinforcement learning. Our automatic and manual evaluations demonstrate that these models are significantly better at generating stories that follow the desired emotion arcs compared to baseline methods, without sacrificing story quality.

1. Introduction

While neural language models have become remarkably good at generating fluent text, they often struggle with the "soul" of storytelling: the emotional journey. Most existing automated storytellers focus on plot consistency or topic adherence but treat emotions as an afterthought or a global "sentiment" (e.g., just making a story "happy").

In reality, compelling narratives are defined by change. Cognitive science suggests that readers stay engaged by monitoring the protagonist's shifting internal states. The challenge is that a protagonist's reaction might differ from the general "vibe" of a sentence (e.g., "My boss was angry" is an angry sentence, but the protagonist's reaction is sadness or fear). This paper seeks to bridge this gap by treating the protagonist's emotional arc as a controllable attribute.

2. Methodology

The authors build upon GPT-2 (medium) as their base model and propose three distinct approaches to incorporate emotion:

  1. Emotion Supervision (EmoSup): A supervised learning approach where the desired emotion arc is prepended to the story title as an input prompt (e.g., [joy, anger, sadness] Title: Raw burger).
  2. Emotion-Reinforced (EmoRL) Models: These models use Reinforcement Learning (REINFORCE algorithm) to fine-tune the generator. They use a "mixed loss" function that balances word-prediction accuracy with "Emotion-Consistency" rewards.
The EmoRL framework architecture, showing the interaction between the story generator agent and the environment providing emotion-based rewards.

There are two flavors of rewards used in the EmoRL framework:

  • RL-EM (Reward via Commonsense): Uses COMET, a commonsense transformer, to infer the protagonist's likely emotional reaction to generated events. It calculates a reward based on the similarity (modified Levenshtein distance) between these inferred reactions and the target arc.
  • RL-CLF (Reward via Classifier): Uses a BERT-based emotion classifier to calculate the probability that a generated story segment matches the target emotion. This model proved to be the most effective.

To ensure narrative flexibility, the authors included a neutral state in the emotion arcs. This was crucial for handling transition sentences or mundane plot points that don't carry heavy emotional weight, preventing the model from forcing "melodrama" into every line.

3. Experimental Setup

The primary dataset used is ROCStories, containing ~98,000 five-sentence stories. Since ROCStories lacks emotion labels, the authors developed an automated pipeline using coreference resolution to identify the protagonist, COMET to infer reactions (xReact), and the NRC Affect Intensity Lexicon to map these to five basic emotions: anger, fear, joy, sadness, and neutral.

4. Results and Analysis

A key insight of this work is the distinction between the protagonist's emotion and the sentence's general sentiment. By using COMET's relations, the model specifically tracks the main character's internal state. While EmoSup only sees the emotion arc at the start, the EmoRL models receive feedback during the generation process, helping the model stay on its emotional tracks.

4.1 Automatic Evaluation

In automatic evaluations, the RL-CLF model significantly outperformed baselines (like Fusion and Plan&Write) in "Arc-accuracy"—the ability to hit the requested emotional beats. Remarkably, the models improved emotional adherence without increasing perplexity or reducing the diversity of the text. In fact, the RL models often produced more diverse stories (higher "Distinct-n" scores) than the supervised baseline.

Quantitative comparison of models across emotion-specific metrics, highlighting the superior performance of RL-based models in arc accuracy and consistency.

4.2 Human Evaluation

In manual A/B testing on Amazon Mechanical Turk, humans preferred the stories generated by RL-CLF over 60% of the time compared to uncontrolled GPT-2, confirming that the model generates more engaging and faithful narratives.

5. Conclusion

This work introduces the first neural storytelling framework designed to track and control the emotional trajectory of a story's protagonist. While effective, limitations include the "discourse-agnostic" nature of current commonsense models and a focus on a single character. Future work could explore more dynamic arcs and complex emotional interactions between multiple characters.

Summary updated

Abstract

Emotions and their evolution play a central role in creating a captivating story. In this paper, we present the first study on modeling the emotional trajectory of the protagonist in neural storytelling. We design methods that generate stories that adhere to given story titles and desired emotion arcs for the protagonist. Our models include Emotion Supervision (EmoSup) and two Emotion-Reinforced (EmoRL) models. The EmoRL models use special rewards designed to regularize the story generation process through reinforcement learning. Our automatic and manual evaluations demonstrate that these models are significantly better at generating stories that follow the desired emotion arcs compared to baseline methods, without sacrificing story quality.
BibTeX
@inproceedings{brahman2020modeling,
  author = {Faeze Brahman and Snigdha Chaturvedi},
  title = {{Modeling Protagonist Emotions for Emotion-Aware Storytelling}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2020},
  url = {https://openreview.net/forum?id=_JJGoMQtfDg},
}

Loading related content...

Figure