Paper Session #3: Virtual Characters and Agents
Talk 1: A Knowledge-Based Framework for the Collaborative Improvisation of Scene Introductions
How do improvers do their stuff? i.e. reach shared understanding, reason about the scene, co-create a narrative?
Built a game called Party Quirks
Three Line Scene? Tiny West?
Talk 2: A New Approach to Social Behaviour Simulation: the Mask Model
Social performance – different roles in different contexts
Each mask contains a level of influence from 0 to 1, for example, religion as a mask – a low value would describe a moderate believer, while a high value would correspond to a fanatic.
If a character has a certain role, when a pre-condition is triggered, an action on an entity is suggested with a certain strength, for example, if I come here to present, I could get up here and be silent, but the strength of the pressure to present is quite high.
It is also important to note that the character will usually have access to the rules denoting other character roles, even without possessing those masks themselves, e.g. we all know we should not take other people’s computers, even though few of us may actually be thieves.
There are three layers to each character – self-perception layer (can model self-deception) is always active, social layer (including religion, politics, culture, etc) which includes beliefs and their respect for social constraints, and the interpersonal layer (persistent, or only active in the presence of a certain character, e.g. mother, girlfriend or boss).
Need to manage the conflicts between the masks – e.g. Romeo’s different masks say he should love Juliet, but his other mask that he shows toward his parents says no. How do we decide how he will act in a given situation?
What is the perceived possibility of him being in a given social situation where this conflict will arise? What are the pros and cons? What is the interpreted morality of each action, and which one better suits with the needs of the character?
Ego-esteem – desire to do the thing, moral soundness of the thing, possibility of getting caught
When two opposing factors result in one of the factors being subverted because the other is far stronger, this becomes a dramatic turn in the plot.
It’s more efficient to create stereotypical masks, then assign several of those to each character, rather than custom-designing each mask for each character.
This could also be used to create personalized NPCs, based on interpreting the moral actions of the PC within the culture of the simulated world.
Murray – “This is akin to modeling all the muscle systems of a dog so you can get an accurate representation of a dog’s choice of actions, but the cartoon dog, that behaves like a dog, is far more convincing. Where is the value for the interactor in including all of these extra systems?”
Talk 3: Perceived or Not Perceived: Film Character Models for Expressive Natural Language Generation
Dialogue generation – I want to flatter you, I’m a friendly person, but I’m feeling hesitant. Can you generate dialogue that reflects all of that?
Deconstructing elements of language to provide subtext – long or short sentences meaning terseness or pontification, use of larger words indicating education, etc.
More parameters = better dialogue?
… Why try to generate natural dialogue by creating caricatures of particular stereotypes and applying them to generic dialogue? Someone has to write that generic dialogue in the first place.
Talk 4: Representing Dramatic Features of Stories through an Ontological Model
Extracting generic information from video clips – i.e. actions, characters, motivations, content
Automatic segmentation of narrative units? Automatic storyboarding and previs?
DMO – Dramatic Media Object?