Perceived or Not Perceived: Film Character Models for Expressive NLG

This paper presents a method for learning models of character linguistic style from a corpus of film dialogues and tests the method in a perceptual experiment. We apply our method in the context of SpyFeet, a prototype role playing game. In previous work, we used the

Personage

engine to produce restaurant recommendations that varied according to the speaker’s personality. Here we show for the first time that: (1) our expressive generation engine can operate on content from the story structures of an RPG; (2)

Personage

parameter models can be learned from film dialogue; (3)

Personage

rule-based models for extraversion and neuroticism are be perceived as intended in a new domain (SpyFeet character utterances); and (4) that the parameter models learned from film dialogue are generally perceived as being similar to the character that the model is based on. This is the first step of our long term goal to create off-the-shelf tools to support authors in the creation of interesting dramatic characters and dialogue partners, for a broad range of types of interactive stories and role playing games.