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.

Introduction

While large-scale language models like GPT-2 can generate fluent text, they often struggle with long-term coherence and lack the ability to follow specific narrative constraints. Specifically, in storytelling, the emotional journey of the protagonist is a key element that keeps readers engaged. Existing "emotion-aware" models usually focus on the overall sentiment of a story or just the ending. However, a story's general sentiment (e.g., "The storm was terrifying") is often different from the protagonist's internal state (e.g., the protagonist feeling "determined" to survive). This paper addresses the challenge of controlling the protagonist's specific emotional evolution throughout the beginning, middle, and end of a generated story.

Methods

We utilize GPT-2 (medium) as the backbone and propose three distinct models to control emotional trajectories:

  1. EmoSup (Emotion Supervision): A conditional training approach where the desired emotion arc is prepended to the story title as an input (e.g., [joy, anger, sadness] Title: Raw burger).
  2. EmoRL (Emotion-Reinforced) Models: These use Reinforcement Learning (REINFORCE algorithm) to fine-tune the model using a "mixed loss" that balances language modeling (fluency) with emotional faithfulness. They introduce two novel reward functions:
  3. EC-EM (Emotion Consistency via Commonsense): Uses COMET (a commonsense transformer) to infer how a protagonist would react to events in the generated text. It calculates a modified Levenshtein distance between these inferred reactions and the target emotion arc.
  4. EC-CLF (Emotion Consistency via Classifier): Uses a BERT-based emotion classifier to predict the probability that each segment of the generated story matches the target emotion.
Architecture of the EmoRL framework featuring the story generator and reward mechanisms.

Datasets and Annotation Pipeline

We use the ROCStories corpus (~98,000 five-sentence stories), segmented into a 1:3:1 ratio (beginning, body, end) to define the arc. Since ROCStories lacks emotion labels, we built an automated pipeline using coreference resolution to find the protagonist and COMET to label their emotional state sentence-by-sentence. The NRC Affect Intensity Lexicon is then used to map open-ended phrases from COMET into five basic emotion categories: anger, fear, joy, sadness, and neutral.

Key Results

Our automatic and manual evaluations demonstrate that modeling the protagonist's internal state leads to more structured and human-like narratives.

  • Faithfulness: The RL-CLF model significantly outperformed all baselines (including Fusion and Plan&Write) in "Arc-acc"—the ability to correctly hit all three stages of the requested emotion arc.
  • Fluency: Interestingly, the proposed models outperformed a standard fine-tuned GPT-2 on perplexity (PPL), suggesting that having an emotional "plan" helps the model maintain better linguistic focus.
  • Human Evaluation: In pairwise comparisons, human judges preferred the RL-CLF stories over 60% of the time compared to uncontrolled GPT-2 models, citing both better emotional adherence and better content quality.
Automatic evaluation results comparing the proposed models against baselines across various metrics.

Discussion and Limitations

The paper emphasizes that the protagonist's internal state is the "engine" of the narrative. Using commonsense as a reward (via COMET) allows the model to understand that "Tom yelled at the waiter" implies anger even without explicit emotion words. However, some limitations remain:

  • Discourse Agnosticism: The COMET model evaluates sentences in isolation, sometimes missing context from previous sentences.
  • Single Protagonist Focus: The current work only tracks the most frequent character, whereas real stories often involve complex emotional interactions between multiple characters.
  • Fixed Arc Length: The model is currently constrained to a three-stage arc (beginning, body, end).

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},
}
Figure