A Survey of Listener Behavior and Listener Models for Embodied Conversational Agents - Coverbal Synchrony in Human-Machine Interaction - page 250

before (Gratch et al., 2002; Cassell and Thórisson, 1999; Granström et
al., 2002); however, nobody tried to associate one or more meanings
to backchannel signals. Moreover, most of the previous experiments
consisted in evaluating the listener's signals during human-human
interactions rather than in human-agent interaction.
3.1 Non-verbal signals
In two perceptual studies (Bevacqua et al., 2007; Heylen et al., 2007),
Bevacqua, Heylen and colleagues asked people to watch a set of
backchannel signals performed by a virtual agent and to assign a
meaning to each of them. The meanings were chosen among the
communicative functions suggested in the literature (Allwood et al.,
1993; Poggi, 2003). All backchannel signals were displayed context-
free , that is, the discursive context of the speaker's speech was not
defined. The first experiment presented in Bevacqua et al. (2007)
served as a pilot test to understand how users perceived backchannel
signals displayed by a virtual agent. Were participants able to assign
a meaning to these signals? Could they easily distinguish between
positive and negative signals? Did backchannels shown by an ECA
allow polysemy and synonymy? For polysemy and synonymy, the
authors meant that the same signal could have different meanings
and a single meaning could be expressed with different signals or
combination of signals.
Fourteen video clips were created using signals chosen among
those proposed in the literature (Allwood and Cerrato, 2003; Poggi,
2007), for example head nod , head nod+smile , head nod+raise of the
eyebrows . Some signals were simple, containing just a single action
(like head nod or head shake ), while others were obtained by combining
several actions (like head shake+frown+tension of the lips ). To display
the signals, the 3D agent Greta (Niewiadomski et al., 2009) was
used. Twelve French participants took part in the evaluation. They
watched one video at a time and then selected for each of them
none, one or more meanings among the 12 proposed: agreement,
disagreement, acceptance, refusal, interest, not interest, belief, disbelief,
understanding, not understanding, liking, not liking.
To analyze the data, the authors coded the answers given by the
subjects as positive or negative, according to the following principles:
agreement, acceptance, liking, belief, understanding and interest
were considered as positive answers and disagreement, refusal,
not liking, not understanding and not interest were considered as
negative answers. Results confirmed that participants were able to
assign a meaning to backchannel signals even when displayed by a