In traditional computational linguistics, a character is often treated as a fixed point in a high-dimensional space. Whether through a single latent distribution Bamman et al. or a static embedding vector Inoue et al., these models assume that an entity’s identity is immutable. However, this "static portrait" approach fails to capture the very essence of storytelling: the character arc.
Recent research argues that to truly understand or simulate a character, we must move away from fixed representations and toward dynamic models that evolve alongside the narrative.
The Limitation of the Static Average
Many current natural language processing approaches create character representations by averaging features across an entire text. As Anderson notes, this methodology creates a "static portrait" that ignores character "roundness" and growth. By collapsing a character's journey into a single average score, models lose the ability to distinguish between a protagonist's initial state and their final transformation.
This limitation is not just theoretical; it has practical consequences for model performance:
- Loss of Narrative Arcs: Rosenbusch argues that character growth and change are often more critical to audience enjoyment than a fixed set of personality traits, yet quantitative methods often fail to capture this evolution.
- Temporal Hallucinations: In role-playing agents, Park et al. observe that static retrieval methods lead to agents exhibiting knowledge or behaviors from the wrong point in a story’s timeline.
- Rigid Role Assumptions: Jahan et al. point out that assuming a character occupies a single role (like "Hero") prevents models from capturing complex transitions, such as a helper becoming an antagonist.
Characters as Dynamic State Machines
To solve these issues, researchers are reframing character identity as a dynamic state machine. Instead of a fixed vector, a character is represented as a sequence of states that update in response to narrative events.
Chen et al. propose using temporal annotations to update character attributes, skills, and social ties as a story progresses. This allows a dialogue agent to reflect the specific knowledge and emotional state a character would have at a precise moment in the timeline. Similarly, Park et al. introduced a framework that updates character traits chapter-by-chapter, mimicking the way human memory consolidates new experiences into an evolving persona.
Mechanisms of Evolution
How exactly does a model "evolve" a character? Several mechanisms have been proposed to capture different facets of change:
Experience-Based Sequences
Rather than treating personality as a tag, Jiang et al. model it as a chronological sequence of extracted experiences. By treating the interaction between personality and context like a chemical "ion exchange," the authors generate story endings that are rational consequences of a character's established history.
Moral and Emotional Trajectories
Character development often hinges on internal shifts. Li et al. developed the CharMoral dataset to track how a character's moral stance fluctuates based on their intentions and the consequences of their actions. For emotional consistency, Wei et al. use Markov chains to ensure that emotional shifts are psychologically plausible rather than erratic, allowing for realistic character growth.
Relational Dynamics
Identity is often situational. Sang et al. highlight that characters wear different "identity masks" depending on who they are talking to. This suggests that a dynamic representation must also account for the evolving social network of the narrative, a concept furthered by Perri et al., who use Graph Neural Networks to capture the shifting relational topology of complex literary worlds.
Toward Realistic Simulation
The shift toward dynamic modeling is particularly vital for simulating human behavior over long periods. In the context of mental health, Wang et al. demonstrate that simulating a patient's understanding of their condition as an evolving sequence of stages—rather than a static list of symptoms—creates more authentic long-term simulations.
Ultimately, as Abdulhai et al. argue, persona simulation should not be a rigid set of rules. True consistency in a character comes not from staying the same, but from evolving in a way that is justified by their experiences. By embracing temporal evolution, computational models can finally move past the "static portrait" and begin to capture the narrative arcs that define human storytelling.
Key Foundational Work
Representing Movie Characters in Dialogues (2019) — contributed 2 concept elements to this direction.
A Bayesian Mixed Effects Model of Literary Character (2014) — contributed 2 concept elements to this direction.
Personality Traits Recognition in Literary Texts (2019) — contributed 1 concept elements to this direction.