DYPLODOC: Dynamic Plots for Document Classification

Anastasia Malysheva, Alexey Tikhonov, Ivan P. Yamshchikov

TL;DR

DYPLODOC introduces a novel method for extracting "narrative arcs"—the emotional and thematic trajectory of a story—to improve document classification. By mapping TV show episode descriptions to a dynamic sequence of genres and emotions, the researchers created a tool that can cluster thousands of shows based on their structural "DNA" rather than just static keywords.

What Problem Does This Paper Address?

While Natural Language Processing (NLP) has become excellent at generating short snippets of text (like chat responses or poetry), long-form narrative generation remains a "bottleneck." The core issue is that long stories require a coherent narrative arc—a structured progression of events and emotions.

Existing attempts to track these arcs often rely on topic modeling, which is notoriously difficult to generalize. For example, a detective story set in Victorian England and one set on Mars might have completely different "topics" (carriages vs. spaceships), making it hard for a computer to recognize they share the same narrative structure. The authors seek a more abstract, universal way to represent how a plot evolves over time.

Methods

The authors developed the Plot Dynamics Extraction (PDE) Tool. Instead of looking for specific keywords, the tool treats a story (specifically a TV season) as a sequence of states.

  1. Dual Classifiers: They use two pre-trained models to "score" every episode description:
  2. Genre Classifier: Maps text to a 20-dimensional space of movie genres (e.g., Action, Comedy, Horror).
  3. DeepMoji: Maps text to an 11-dimensional space of emotional labels (e.g., Love, Fear, Mad, Sad).
  4. The Narrative Trajectory: Every season is represented as a path through these 31 dimensions (a "union of two simplexes"). To prevent the math from breaking when a probability hits zero, they implement custom smoothing procedures.
  5. Feature Engineering: To turn these "paths" into data a machine can classify, they calculate synthetic features for each dimension, including the mean, median, variance, and the "area under the curve" for each emotion and genre throughout the season.
  6. Clustering: Finally, they use K-means clustering and Principal Component Analysis (PCA) to group seasons with similar dynamic signatures.

Datasets

The researchers introduced the DYPLODOC dataset, a massive collection of TV metadata scraped via the TVmaze API.

  • Scale: Includes 13,813 TV shows and 21,962 seasons.
  • Granularity: Contains over 300,000 individual episode descriptions.
  • Content: Each entry includes the main storyline, character info (via Named Entity Recognition), and genre tags.

Key Insights and Ideas

  • Abstraction is Key: By projecting specific plots onto a fixed "state space" of genres and emotions, the researchers created a "universal language" for narratives. This allows them to compare a sitcom to a dark dramedy based on how their "sadness" or "humor" levels fluctuate over 10 episodes.
  • The Power of "And": One of the biggest "aha" moments was finding that combining genre dynamics with emotional (emoji) dynamics yielded significantly better results than using either alone. The "what" of the plot and the "feel" of the plot are complementary signals.
  • Structural DNA: The paper suggests that a show's genre isn't just a label; it's a pattern of movement. A "Dark Comedy" isn't just funny and sad at the same time; it's a specific way of oscillating between those two states.

Key Results

  • Classification Accuracy: The researchers measured "Mutual Information" (MI) to see how well their unsupervised clusters matched human-assigned genre tags. The combined dynamic features achieved an MI of 0.57, significantly higher than using genre (0.47) or emoji (0.36) dynamics in isolation.
  • Case Study (BoJack Horseman): The tool successfully mapped the fourth season of BoJack Horseman, showing a clear dominance of "Comedy" and "MAD" (anger) throughout, but correctly identifying a spike in "SAD" and "Drama" during a specific narrative arc involving the protagonist's mother in a nursing home.
  • Cluster Purity: In their t-SNE visualizations, many clusters showed high "dominant tag" percentages (up to 100%), proving that plot dynamics are a highly effective proxy for traditional genre classification.

Limitations and Future Work

  • Data Gaps: About 36% of the seasons in the original dataset lacked human-assigned tags, forcing the researchers to exclude them from the validation phase.
  • Classifier Dependency: The system is only as good as its underlying classifiers. If the Movie Genre or DeepMoji models misinterpret a nuanced episode description, the entire narrative arc for that season becomes skewed.
  • Future Direction: The authors suggest using these extracted arcs as "semi-supervision" for generative models. Instead of telling an AI to "write a story," you could give it a specific "narrative arc" (e.g., "start happy, get progressively angrier, end with a joke") to guide the generation of longer, more coherent texts.

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},
}

Loading related content...

Figure