Information Technology Reference
In-Depth Information
the latter is linked to another computer controlled by the wizard, who can
answer in written or oral form, by generating his/her own utterances or
choosing from some predefined patterns. If the simulation is correctly carried
out, the subject does not discover the deception and adopts the behavior
he/she would have with a machine, which already allows us to analyze this
type of behavior. The dialogues are recorded to allow the designers to detect
problematic situations, the cases in which the wizard's reactions disturbed the
subject and obviously the cases of misunderstandings. By focusing on the
incriminated modules, the designers can thus improve their system and make
it more robust. To detect the moments where a subject is disturbed in a
reliable manner, and to detect the moments of incomprehension or those of
hesitation, additional techniques are implemented at the same time as the
dialogue recording: we can, for example, track the subject's face, so as to
record his/her facial emotions or use an eye-tracker to detect his/her eye
movements and deduce the observations on his/her attention and, when there
is a shared visual scene, on the objects observed, for example, when solving a
reference.
For a Wizard of Oz to be usable, it requires the experimental conditions to
be well defined. However, there are as many ways to design a Wizard of Oz as
there are systems: the wizard can be one of the designers, which allows
him/her to generate a behavior close to the one of the targeted system, but
he/she can also be a second experiment subject, who does not know the goal
of the experiment and only applies a set of rules aiming to simulate the
system (double-blind principle, also known as ghost in the machine).
Moreover, the subject and the wizard's messages go through a computer
system, and the latter can integrate a specific process so as to complicate the
experiment. In [RIE 11], noise is added to the subject's utterances before
transmitting them to the wizard, which disturbs the dialogue and allows us to
increase the frequency of clarification requests. In this case, one word was
randomly deleted, but we could very well imagine that a word would be
replaced by another or by other local heuristics, and not just for the subject's
utterances but also for the wizard. Obviously, it is easier to carry this out in a
written dialogue than in an oral dialogue. In the case of the very complex
methodology implemented in [RIE 11], the Wizard of Oz's goal is fourfold:
to observe dialogue situations; to create a study corpus and thus model a
machine learning model; to create a learning corpus; and to contribute to a
user model specification, that is a computer program whose goal is to
simulate the behavior of a system's user. Each step and goal is accompanied
with precautions, assessments and confrontations with other methods such as
Search WWH ::
Custom Search