DYPLODOC: Dynamic Plots for Document Classification

Anastasia Malysheva, Alexey Tikhonov, Ivan P. Yamshchikov

Abstract

Narrative generation and analysis are still on the fringe of modern natural language processing yet are crucial in a variety of applications. This paper proposes a feature extraction method for plot dynamics. We present a dataset that consists of the plot descriptions for thirteen thousand TV shows alongside meta-information on their genres and dynamic plots extracted from them. We validate the proposed tool for plot dynamics extraction and discuss possible applications of this method to the tasks of narrative analysis and generation.

Introduction

While Natural Language Processing (NLP) has become highly effective at generating short snippets of text, long-form narrative generation remains a significant bottleneck. The core challenge lies in maintaining a coherent narrative arc—the structured progression of events and emotions over time. Existing methods often rely on topic modeling, which fails to generalize across different settings (e.g., a Victorian mystery versus a sci-fi mystery). DYPLODOC addresses this by proposing an abstract, universal representation of how a plot evolves, treating a story's structure as its "DNA."

Methodology: Plot Dynamics Extraction (PDE)

The authors developed the Plot Dynamics Extraction (PDE) tool to treat a story as a sequence of states rather than a collection of keywords. The process involves:

  1. Dual Classifiers: Episode descriptions are processed through a Genre Classifier (20 dimensions) and DeepMoji (11 emotional dimensions, such as Love, Fear, Mad, and Sad).
  2. Narrative Trajectory: Each season is mapped as a path through this 31-dimensional space. Custom smoothing procedures are applied to handle zero-probability states and ensure mathematical consistency.
  3. Feature Engineering: Synthetic features—including mean, median, variance, and area under the curve—are calculated for each dimension to represent the season's dynamic signature.
  4. Clustering: K-means and Principal Component Analysis (PCA) are used to group seasons with similar structural patterns.

The DYPLODOC Dataset

The researchers introduced a massive collection of TV metadata scraped via the TVmaze API, comprising:

  • 13,813 TV shows and 21,962 seasons.
  • Over 300,000 individual episode descriptions.
  • Metadata including main storylines, character information (via Named Entity Recognition), and genre tags.

Results and Discussion

The study found that combining genre dynamics with emotional dynamics yielded superior results compared to using either in isolation. The combined features achieved a Mutual Information (MI) score of 0.57, significantly outperforming genre-only (0.47) and emoji-only (0.36) models. This suggests that the "what" of the plot and the "feel" of the plot are complementary signals.

t-SNE visualization of TV show clusters based on dynamic plot features, showing high alignment with traditional genre tags.

A case study of BoJack Horseman Season 4 demonstrates the tool's precision. The model captures the show's baseline dominance of "Comedy" and "MAD" (anger) while correctly identifying a significant spike in "SAD" and "Drama" during a specific narrative arc involving the protagonist's family history in the middle of the season.

Narrative arc of BoJack Horseman Season 4 showing fluctuations in genre and emotion probabilities throughout the season.

The t-SNE visualizations confirm that these dynamic clusters often align closely with human-assigned genre tags, with some clusters reaching 100% purity. This proves that plot dynamics are a highly effective proxy for traditional genre classification.

Conclusion and Future Work

DYPLODOC demonstrates that a show's genre is not just a static label but a pattern of movement between emotional and thematic states. While the system is currently dependent on the accuracy of its underlying classifiers and faces some data gaps (36% of seasons in the raw data lacked tags), it opens new doors for narrative generation. The authors suggest using these extracted arcs as "semi-supervision" to guide AI in generating longer, more structurally coherent texts by providing a specific emotional trajectory to follow.

Summary updated

Abstract

Narrative generation and analysis are still on the fringe of modern natural language processing yet are crucial in a variety of applications. This paper proposes a feature extraction method for plot dynamics. We present a dataset that consists of the plot descriptions for thirteen thousand TV shows alongside meta-information on their genres and dynamic plots extracted from them. We validate the proposed tool for plot dynamics extraction and discuss possible applications of this method to the tasks of narrative analysis and generation.
BibTeX
@inproceedings{malysheva2021dyplodoc,
  author = {Anastasia Malysheva and Alexey Tikhonov and Ivan P. Yamshchikov},
  title = {{DYPLODOC: Dynamic Plots for Document Classification}},
  booktitle = {Mmbd/mlis},
  year = {2021},
  doi = {10.3233/faia210283},
}
Figure