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.
The Challenge of Inferring Persona from Dialogue
While personality is often quantified using the "Big Five" traits in psychology, storytelling relies on "tropes"—social roles like the "Hardboiled Detective" or the "Lovable Rogue." Previous computational efforts to categorize these personas often relied on high-level Wikipedia plot summaries. However, these summaries are third-party overviews rather than direct character expressions.
Researchers from the MIT Media Lab argue that for better conversational agents and narrative analysis, we must infer personas directly from dialogue. The primary challenge is that dialogue is inherently noisy; characters spend much of their time on generic "chit-chat" (e.g., asking for the time) that does not reflect their underlying persona. The goal of this work is to filter the "signal" of personality from the "noise" of everyday conversation.
Architecture: Attentive Memory Networks
The proposed architecture, the Attentive Memory Network, addresses the sparsity of persona through a two-stage process designed to aggregate information across multiple dialogue snippets.
Multi-Level Attentive Encoders
- Word-level Attention: A GRU-based encoder processes individual snippets, using self-attention to identify words most indicative of a trope (e.g., focusing on aggressive language in a villainous persona).
- Snippet-level Attention: Because a single quote is rarely enough to define a persona, the model processes a batch of $N$ snippets. It learns to assign higher weights to "trope-reflective" snippets while ignoring generic dialogue.
Knowledge-Store Memory Module
The model incorporates domain knowledge by scraping descriptions of 72 tropes from TVTropes.org. These descriptions are converted into "skip-thought" vectors to initialize a memory module. During training, a triplet loss function forces dialogue-based embeddings to align with their corresponding trope descriptions in the embedding space.
Filtering Signal from Noise
A key insight of the study is the "sparsity" of persona. The success of the snippet-level attention mechanism confirms that persona-defining traits are not present in every sentence. By allowing the model to selectively attend to specific parts of a script, classification accuracy improved from a baseline of ~28% to over 60%.
As shown in the attention visualization, highly specific or dramatic lines (e.g., "Rumors of my death have been greatly exaggerated") receive significantly higher weights than shorter, more generic utterances. Furthermore, the researchers found that including "3-way data"—the character's lines, the interlocutor's response, and scene context (like [shoots him])—dramatically improved performance, as personas are often defined by how they interact with others.
Key Results and Narrative Analysis
The best-performing model reached 75% accuracy (F1 score of 0.75), a significant leap over single-snippet GRU baselines. Performance scaled positively with the number of dialogue snippets ($N_{diag}$), proving the model's ability to aggregate character history effectively.
Beyond classification, the authors used these embeddings to cluster movies based on their character distributions. The model successfully grouped "Gangster" films (e.g., Scarface, Goodfellas) and "Dark/Psychological" films (e.g., Taxi Driver, Inception) together, demonstrating its utility for automated narrative and genre analysis.
Conclusion and Future Work
By combining multi-level attention with external knowledge stores, this research provides a robust framework for extracting social roles from messy, real-world dialogue. Future applications include "personality-infused" dialogue agents that can maintain consistent, trope-specific personas, moving beyond the generic responses of current conversational AI.
Reference: Chu, E., Vijayaraghavan, P., & Roy, D. (2018). Learning Personas from Dialogue with Attentive Memory Networks. Proceedings of EMNLP 2018.
Summary updated