Co-speech Gesture Generation for Embodied Agents and its Effects on User Evaluation - Coverbal Synchrony in Human-Machine Interaction - page 234

competence and believability, whereby nonverbal behavior consisted
of facial expressions accompanied by emotional gestures. It was found
that all dependent measures (warmth, competence and believability)
increased with the number of modalities used by the agent.
5.1 Effects of gesture generation with
different GNetIc models
The GNetIc account has been evaluated with regard to whether the
generated gestures can be beneficial for human-agent interaction
(Bergmann et al., 2010). Two aspects were addressed in this study.
First, the quality of the produced iconic gestures as rated by human
users, and second, whether an agent's gesturing behavior could
systematically alter a user's perception of the agent's warmth and
competence as the two fundamental dimensions of social cognition.
The warmth dimension captures whether we judge someone as are
friendly and well-intentioned, and the competence dimension captures
whether someone has the ability to deliver on those intentions.
A number of studies have shown that warmth and competence
assessments determine whether and how we intend to interact with
others (Cuddy et al., 2011): We seek the company of people who are
assumed to be warm and avoid those who appear less sociable (i.e.
cold). With regard to competence, we prefer to cooperate with people
we judge as competent, while incompetent people are disregarded.
To investigate both questions, the flexibility afforded by GNetIc
to generate speech-accompanying gestures was exploited in a study
comparing five different conditions: two individual conditions (ind-1
and ind-2 with GNetIc networks learned from the data of individual
speakers: subject P5 in ind-1, subject P7 in ind-2), a combined
condition with a network generated from the aggregated data of five
different speakers (including P5 and P7), two control conditions (no
gestures at all and simple random choices at the chance nodes in the
network). Note that in all conditions, the gestures were produced from
identical input values and accompanied identical verbal utterances.
The random condition resulted still in appropriate iconic gestures,
as the decision nodes still applied their sound mappings, but these
gestures could occur at atypical positions in the sentence or employ
random handedness.
In a between-subject design, 110 participants received a description
of a church by the embodied agent Max, produced fully autonomously
with a speech and gesture production architecture containing GNetIc.
Immediately after receiving the descriptions, participants filled out