Information Technology Reference
In-Depth Information
precision level, a classic method for all the NLP systems [CHA 04], adapted
to the separate assessment of modules but not the system as a whole.
Finally, the system obtained undergoes user tests, i.e. subjects are once
more involved, who are neither the designers nor the original Wizard of Oz
subjects, and their opinions are collected. They are analyzed by the designers,
who then decide what improvements should be made, and might carry out a
new development phase: improving the core of the system, then the
understanding and generation of modules and finally the voice recognition
and text to speech.
This scenario covers the weak points of the scenario discussed in
section 3.1.1. However, even with a rational architecture, most of the
processes are chain processes, and the essential aspects of the oral language
such as prosody are ignored. This is typically the point when we consider the
issues described previously on the consequences of not taking the oral
language's morphology into account.
3.1.3. A scenario today
A design scenario today would have to provide answers for all the weak
points mentioned previously, and all the issues of section 1.3. Without
discussing them again, let us note that a system's development can now
require not just the existing modules, but a toolkit that provides a completely
configurable general framework. As for any framework, it could become
penalizing due to the limits it imposes. In this case, the development of
additional modules is crucial, and is the focus of previous experiments, tests
and assessments through metrics that are more precise than recall and
precision calculations. The MMD system itself can now be globally assessed
with methods that calculate the task's completion level, on indexes such as the
number of speech turns to reach the originally defined goal. We will discuss
these aspects in Chapter 10. Moreover, all the development stages are being
rationalized, so that nothing is done blindly or by chance. For example, the
corpuses are made for each of the identified modules, and these corpuses are
themselves broken up into various parts: one part for training (machine
learning is integrated into the stages that need it most), and one part for tests.
Another example, the Wizard of Oz, is carried out with additional constraints
and precautions. This is the type of constraints we will now detail, going over
a few emblematic design stages.
Search WWH ::




Custom Search