The Development Stages of a Dialogue System - Man-Machine Dialogue: Design and Challenges - page 59

finally, the implementation creates no surprises for designers who are not at
their first MMD. In any case, the clarity of specification, the verification of
available resources and the exchange between various specialists are classic
solutions that may not be original but are essential in designing the system.
Is there an order in which to develop the modules? It appears at first that if
theinput and output of each module wascorrectly defined and remained stable,
the order does not matter and the development can be split among various
people. In reality, as we showed with the scenarios in section 3.1, it is better
to start with the system's core, that is the dialogue manager, and end with the
surroundingmodulessuchasspeechrecognitionandtexttospeech.Takingthis
precaution will provide more tolerance for the specification's little mistakes
and to avoid surprises during the development phase.
Is there a list indicating resources? It obviously all depends on the
processing abilities considered by the system, and the possibilities of online or
offline retrieval of existing resources. With what we have already presented, as
well as a few additional elements that can be deduced and that we will describe
later, here is an indicative list that shows the diversity of models involved in an
MMD system:
- Domain models: database of instanced objects, potentially visible and
referable, for example SHRDLU cubes and pyramids, and the physical model
describing potential behaviors and evolutions of these objects.
- MMI models: in the case of an MMD with a supplementary interface
displayed on screen, an operation model of the latter (actions-reactions and
rules of priority of MMI commands over voice commands).
- Design models: ontology, graph or tree of concepts in question, with both
objects and actions that can be carried out on these objects.
- Task models: sequence of actions leading to the task's completion,
conditions and decision tree;
- Acoustic-phonetic models: for speech recognition, as well as for text to
speech (they are not the same models but can have some parts in common).
- Prosodic models: tone outlines, rules for putting in relief, for analysis as
well as generation (same comment).
- Symbolic linguistic models: lexicon, grammars, semantic structures,
types of logical forms, rules for enriching these logics, deduction or induction
mechanisms, etc.