The Development Stages of a Dialogue System - Man-Machine Dialogue: Design and Challenges - page 58

system's many modules, he/she has to ignore certain phenomena, such as all
the associative anaphora, to achieve an operational resolution system in time
and with reasonable means. It is a diminishing factor (and hard to admit), but
it happens.
As another example of simplification linked to multimodal dialogue, let us
cite the processing of combined multimodal references such as a single
demonstrative gesture linked to two or three referential expressions, or several
gestures linked to the same referential expression (or to several referential
expressions but without matching each to its own). It is very difficult to
design a multimodal fusion module able to identify these situations, and it
quickly becomes time-consuming with the inevitable technical problems such
as detecting the beginning and end of a multimodal utterance, the recording of
the gesture's trajectory and the temporal synchronization management. So as
to develop a system able to have reaction times close to human reaction times,
the low-level input processing must be very fast and avoid complex
management of all the gestures - referential expressions matching hypotheses
at the end of utterances. Thus, and this is the case for many systems, we tend
to forget about this type of situation and focus on demonstrative references
involving a single gesture and a single referential expression, which is already
complicated enough (see Chapter 6). In the end, the system only processes a
subset of the phenomena identified at first, but at least it works. Of course, in
the case of system designed by a full team of developers, or in the case of an
existing system being reused, the problems do not appear in the same way. In
those two cases, the process specification of the processing is directly linked
to the system architecture's definition (see Chapter 4).
3.2.5. Resource writing and development
The real issues appear during implementation, as it usually is in NLP.
When programming the main algorithm of a module, we notice that there are
not enough resources, that carrying out the process is more complex than
predicted and that it is necessary to reduce the number of phenomena
processed. We might also notice that there is a piece of missing input data for
the module, for example a prosodic aspect that we had neglected but that ends
up an important setting at a given moment. We could also notice that the
module's output will not be as comprehensive or precise as planned. We could
notice that the algorithm execution is slower than expected (or hoped). And