Co-speech Gesture Generation for Embodied Agents and its Effects on User Evaluation - Coverbal Synchrony in Human-Machine Interaction - page 232

Figure 3. Example gestures simulating different speakers, each of which produced for the
same referent (a round window of a church) in the same initial situation.
hands adopt a C-shape while in the simulation for P8, an O-shape is
used with the right hand only. P1 and P15 both use drawing gestures
which, however, differentiate in their handedness.
4.3 Data-based evaluation
In the following, we will look at whether and how GNetIc models
afford an adequate approximation of empirically observed behavior.
For this purpose, the gestures determined by GNetIc were compared
with those found in the SaGA corpus (Bergmann and Kopp, 2009a;
Bergmann and Kopp, 2010). To this end, the data corpus the GNetIc
models were built from (SaGA data from five different speakers:
473 noun phrases, 288 gestures) has been divided into training data
(80%) and test data (20%). The training set was used for structure
learning and parameter estimation of the decision networks. For each
speaker's test set, it was tested whether the gestures generated with an
aggregated decision network (learned from all of the five speakers) as
well as the gestures generated with the individual decision networks
(learned from only this speaker's data).
In total, individual networks achieved a mean accuracy of 62.4%
(SD = 11.0), while the mean for the aggregated network was 57.8%
(SD = 22.0). By trend, all individual networks performed better than
networks learned from non-speaker specific data. Particularly for the
case of P5, the accuracy of both, technique and handshape decisions,
is remarkably better with the individual network than with the general
one.
In contrast to chance nodes, which are individualized both in their
connections to the input nodes and the local conditional probabilities,
the decision nodes employ identical rules in each network. Yet,
decisions are made in a particular order and decision nodes follow
to operate upon previously determined chance node values. In
consequence, a decision node may sometimes not match the test data
values, because earlier nodes have already produced mismatching
values. Therefore, the rule-based decision nodes were evaluated locally,