Information Technology Reference
In-Depth Information
The point of this method is to better target the diagnostic, since we can identify
whichmonomodalprocessingchainislacking.Theinconveniences, obviously,
are the multiplicity of indispensable reference representations, as well as the
need for a specific assessment method for mutlimodal fusion;
- on the other hand, if we keep on this track, we can consider that the
assessment has to be applied in a modular fashion, that is by using reference
representations to assess the outputs of each of the system's modules. The
main issue with this approach, beyonda strong dependence on the architecture,
is the multiplication of modules and thus that of the matching assessments
and representations: lexical, syntactic and semantic representations of natural
language, facial expressions, gesture trajectories, etc. A modular assessment is
thus an important work load. Another issue: if we multiply local assessments,
it becomes difficult to confront measurements to achieve a global system
assessment. The diagnostic is more precise, but it comes at the expense of
a simple measurement that can help understand the global operation quality.
In addition, and this is a key point, the errors of a module can be made
up by the performances of another module without having any consequence
on the global operation. It is not about compensating the bad running of a
module through an exemplar running of another module, but to compensate
unavoidable errors with relevant making up procedures. The typical example
concerns the speech recognition errors: it is an illusion to hope for a
recognition module that would be able to have perfect performances (100%
of words perfectly recognized) with a very broad vocabulary, or even a few
hundred words. However, if the semantic and pragmatic modules are able to
transform semantic and contextual information as recognition constraints, the
system will be able to find the pronounced utterance in a sure way, even if the
recognition engine's performance is mediocre. In other words, the recognition
module assessment has no point, and only the recognition module and the
semantic and pragmatic modules interaction assessment counts. The modular
assessment is thus not quite as simple to implement nor is it as relevant.
10.2.2. Should a multimodal corpus be managed?
A second issue that arises when implementing an assessment procedure is
the use of a multimodal corpus. In this case, it is a test corpus, which is
obviously different from the corpus previously used to gather phenomena or
for a potential machine learning purpose. The system is input with all the
situations found in the corpus, and we test either the system's output or its
Search WWH ::




Custom Search