Letting Emotions Flow: Success Prediction by Modeling the Flow of Emotions in Books

Suraj Maharjan, Sudipta Kar, Manuel Montes-y-Gómez, Fabio A. González, Thamar Solorio

Introduction

What makes a book a "page-turner"? While literary critics often point to prose style or plot structure, recent research investigates the emotional arc as a primary driver of reader engagement. Previous attempts to predict literary success often relied on "bag-of-words" models or static sentiment analysis, which ignore the sequential nature of storytelling. This paper proposes that the flow of emotions—the way an author moves a reader from hope to despair and back again—is a primary indicator of a book's eventual success.

The Multitask Emotion Flow Model

The researchers developed a deep learning approach that treats a book as a sequence of emotional states. The process begins by dividing each book into $n$ sequential chunks. For every sentence, the NRC Emotion Lexicon is used to count words associated with eight basic emotions (anger, anticipation, joy, trust, disgust, sadness, surprise, fear) plus positive and negative valence.

Each chunk is represented by a 20-dimensional vector containing the mean and standard deviation of these emotion scores. The inclusion of standard deviation captures "emotional volatility" within a chapter. These vectors are fed into a Bidirectional Gated Recurrent Unit (Bi-GRU), allowing the model to capture long-term dependencies by reading the sequence both forward and backward.

Architecture of the Multitask Emotion Flow Model, utilizing Bi-GRU and attention mechanisms to predict success and genre.

An attention layer is added to automatically learn which chunks of the book are most important for the final prediction. Furthermore, the model employs Multitask Learning (MT) to predict both the book's success (Goodreads rating) and its genre simultaneously. This approach uses genre as a regularizer, helping the model understand that a "successful" horror novel has a different emotional flow than a romance.

Data and Evaluation

The study utilized a dataset of 1,003 books from Project Gutenberg across eight genres. Success labels were derived from Goodreads average ratings. Unlike previous studies that only analyzed the beginning of a text, this research evaluated the entire book to determine if the resolution impacts success prediction.

Key Insights: The Importance of the Climax

The attention mechanism revealed a significant pattern: the model places the highest weight on the final two to three chunks of a book. In most genres, the emotional payoff at the climax is the strongest predictor of whether a reader will rate the book highly.

Heatmap of attention weights across book chunks for various genres, showing concentrated importance in the final chapters.

Other key findings include:

  • Volatility Over Average: Features capturing the variation (standard deviation) of emotions were more predictive than average scores. Readers appear to value the contrast between emotions rather than a steady state.
  • Genre Calibration: Multitask learning significantly boosted performance by allowing the model to calibrate expectations based on genre-specific trajectories (e.g., "Riches to Rags" vs. "Man in a Hole").
  • Poetry as an Outlier: Unlike prose, poetry showed emotional weights distributed across the entire text, suggesting a consistent rhythm rather than a singular narrative arc.

Results and Conclusion

The proposed Bi-GRU + Attention + MT model achieved a top score of 69% F1 when using the entire book, significantly outperforming standard SVM baselines (64%) and majority class baselines (50.6%). These results confirm that capturing the full emotional trajectory, especially the shifts in later chapters, is essential for predicting reader reception.

While the study is limited by a relatively small dataset and the word-based nature of the NRC lexicon, it provides a powerful framework for quantifying the "emotional fingerprints" of successful literature. Future work may involve hierarchical models to identify specific suspenseful scenes or help authors refine their storytelling rhythm.

Summary updated

BibTeX
@inproceedings{maharjan2018letting,
  author = {Suraj Maharjan and Sudipta Kar and Manuel Montes-y-Gómez and Fabio A. González and Thamar Solorio},
  title = {{Letting Emotions Flow: Success Prediction by Modeling the Flow of Emotions in Books}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2018},
  doi = {10.18653/v1/n18-2042},
}
Figure