Graphics Reference
In-Depth Information
system that is used to resolve conflicts between behavior suggestions.
This annotation and analysis was critical because existing literature
said little about dynamics of behaviors and further conflict resolution
was to resolve potential conflicts both between the behaviors suggested
by the rules as well as differences across literature sources.
More recently a variety of machine learning techniques have been
explored, including Hidden Markov Models and Latent-Dynamic
Conditional Random Fields to learn the mapping between features of
an utterance and nonverbal behaviors using annotated data face-to-face
interactions. In particular, Lee and Marsella (2010) contrasts several
approaches to learning models of head and eyebrow movement as
well as contrasting the results with the knowledge encoded in NVBG
by the literature approach discussed above.
3.3 Perceptual processing
Perceptual messages are treated differently than generating nonverbal
behavior for the virtual human's utterances. For the perceptual
messages, NVBG is deciding on how to respond to signals about
external events, including the physical behavior of objects, humans or
other virtual humans. These responses, such as looking at a moving
object, can in large measure be reflexive or automatic as opposed to
having an explicit communicative intention like an utterance. Due
to the differences between the perceptual and utterance use cases,
NVBG's perceptual responses analyses use a different processing
pipeline than the utterance handling.
Specifically, NVBG's response is determined by a Perceptual
Analysis stage that leads into the Behavior Analysis and BML
Generation stages discussed previously. The rules used during
Perceptual Analysis take into account what is the perceived behavior
and whether the perceived behavior is above some acceptance
threshold (e.g., an object's speed, size and distance or an event's
duration or magnitude).
3.4 Listener feedback
The listener feedback pipeline handles the virtual human's behavior
while listening to a human or virtual human speaker. The approach
makes a distinction between generic feedback and specific feedback,
handling them using different rule sets. Generic feedback is driven
by speaker behaviors including nods and pauses in speech. Specific
feedback is driven by the virtual human's unfolding interpretation
of, and reaction to, the speaker's utterance, which requires natural
Search WWH ::




Custom Search