Co-speech Gesture Generation for Embodied Agents and its Effects on User Evaluation - Coverbal Synchrony in Human-Machine Interaction - page 226

between the hand configuration and the object shape. Likewise, in
the third speaker's gesture, the round shape of the reference object
is depicted by the movement of the speaker's drawing hand. That is,
there obviously exist several 'techniques' to represent the same kind of
meaning in gestures: whereas in the left hand and the middle examples,
the hands adopt a static posture as a model for the circular shape, the
same shape is depicted by a circular movement of the drawing hand
in the right-hand example. For an adequate account of how meaning
is transformed into gesture form, these representation techniques
certainly have to be taken into account. Concrete mapping rules are
more likely to be found within a set of gestures belonging to the same
representation technique than across all instances of iconic gesture use.
Empirical studies, however, reveal that similarity to the referent cannot
fully account for all occurrences of iconic gesture use (Kopp et al.,
2007). Rather, recent findings actually indicate that a gesture's form is
also influenced by specific contextual constraints such as the discourse
context (Holler and Stevens, 2007; Gerwing and Bavelas, 2004), the
linguistic context (Kita and Özyürek, 2003; Gullberg, 2010; Bavelas et
al., 2002), and gesture history (McNeill, 2005; Bergmann, 2012).
In sum, gesture use in humans is subject to both individual and
common patterns of how meaning is mapped onto gesture form. Thus,
with regard to computational modeling of gesture production, these
complex influences of referent characteristics, contextual factors, and
inter-individual differences should be taken into account. The following
section reviews the state of the art in computational generation models
and Section 4 presents one model in detail—the GNetIc model—that
aims to cover the above-mentioned issues.
3. Generating Co-Speech Gestures for
Embodied Agents
Different modeling approaches have been tested in an effort to
translate systematic characteristics of coverbal gestures, shared among
speakers, into generative models. In line with the two major concerns
in gesture, simulation as suggested by the empirical literature can be
broken down into two categories. First, model-based accounts that focus
on how meaning is mapped onto physical gesture forms as well as
on influences of contextual factors on gesture use. Second, customized
attempts emphasizing inter-individual differences in communicative
behavior trying to model individual gesture style. Finally, there is
another trend in behavior generation, namely data-based attempts to
simulate (individual) speakers' gesturing behavior based on corpora
of gesture use.