Abstract
The ability to infer persona from dialogue can have applications in areas ranging from computational narrative analysis to personalized dialogue generation. We introduce neural models to learn persona embeddings in a supervised character trope classification task. The models encode dialogue snippets from IMDB into representations that can capture the various categories of film characters. The best-performing models use a multi-level attention mechanism over a set of utterances. We also utilize prior knowledge in the form of textual descriptions of the different tropes. We apply the learned embeddings to find similar characters across different movies, and cluster movies according to the distribution of the embeddings. The use of short conversational text as input, and the ability to learn from prior knowledge using memory, suggests these methods could be applied to other domains.
Introduction
In narrative theory, characters often embody recognizable archetypes known as tropes. While humans can intuitively identify a "hardboiled detective" or a "lovable rogue" through a few scenes, automating this classification is a significant challenge for machine learning.
Historically, computational approaches have relied on third-person plot summaries (e.g., Wikipedia) to categorize characters. However, these summaries describe external events rather than the character's intrinsic voice. This research explores whether a persona can be inferred directly from dialogue. This is a "noisy" task, as much of a character's speech is functional exposition or casual filler rather than trope-indicative expression.
Methodology: The Attentive Memory Network
The researchers developed the Attentive Memory Network, a neural architecture designed to filter dialogue noise and leverage domain knowledge. The model consists of three primary components:
- Multi-Level Attention: The model employs word-level attention to identify trope-indicative vocabulary within a sentence and inter-snippet attention to determine which specific quotes in a batch are most revealing of a character's personality.
- Knowledge-Store Memory: This module is initialized with textual descriptions of tropes scraped from TVTropes.org. When processing a new dialogue snippet, the model queries this memory to find the closest archetype match.
- Contextual Encoding: Beyond the character's own lines, the model encodes the "context"—including script action descriptions and the dialogue of interlocutors—to capture social dynamics.
Data and Training
The study utilized three primary datasets:
- CMU Movie Summary Corpus: Provided 72 manually identified tropes across 433 characters.
- IMDB Dialogue Snippet Dataset: A new collection of 13,874 quotes, 70% of which are multi-turn exchanges.
- Character Trope Description Dataset: Detailed textual descriptions of the 72 tropes to serve as prior knowledge for the memory module.
Key Insights
Dialogue as a Sparse Signal
The transition from a standard GRU baseline (~29% accuracy) to the attention-based model (~63% accuracy) highlights that dialogue is a sparse signal. Most speech is "filler"; the attention mechanism is critical for isolating the 10% of lines that actually define a persona.
The Power of Expert Initialization
Initializing the memory module with human-written trope descriptions (Knowledge-Store) proved more effective than letting the model learn a memory from scratch. This suggests that "bottlenecking" a model with high-quality domain knowledge is a powerful strategy for small datasets.
Results and Narrative Analysis
The best-performing model achieved 75.0% accuracy in predicting tropes, a massive improvement over the 28.7% baseline. Beyond classification, the model demonstrated strong qualitative performance in clustering similar characters. For instance, it successfully grouped Carl Fredricksen (Up) and Walt Kowalski (Gran Torino) into a "Grumpy Old Men" category, despite the vast differences in the movies' genres and tones.
Conclusion and Future Work
While the current model is limited by data scarcity (433 characters across 72 tropes), it proves that dialogue is a viable source for persona modeling. Future work may explore translating these methods to real-world personality frameworks like the "Big Five" traits or using these embeddings to generate more consistent, character-driven dialogue for conversational agents.
Summary updated