Reusable System Architectures - Man-Machine Dialogue: Design and Challenges - page 65

parameters allow it. The semantic and pragmatic modules then start to
process the utterance, to enrich it and if the dialogue manager judges it
relevant, it might decide to start talking, with the potential consequence of
interrupting the user if, while the message was generated, he/she is still
talking (and in that case, speech recognition and prosodic and syntactic
analyzers keep on working). Systems, such as Nailon [EDL 05], open up the
way for this type of operation, at least for prosodic aspects. The interesting
factor is the utterance resource, which is constantly evolving as the analyses
are carried out: far from being static data, they become dynamic data,
sometimes vague or incomplete, which system modules will use as they can
without any strict completion or correction (e.g. grammatical) constraints that
are too often imposed. As we can see, the matching run-time architecture
requires serious reflection on the resources and modules involved in the
acquisition of knowledge and the update of the data.
4.1.2. The process flow
The flow of processes carried out by the different architecture modules can
happen in a linear manner, as we have seen, that is as a chain: the output of a
module directly serves as the input for the following module. From a
computational point of view, this is a constraint that can be the source of the
implementation of run-time architecture as it has been defined. It is the
difference between conceptual architecture and its materialization as software
architecture: a chain software architecture only makes sense if the conceptual
architecture itself is a chain architecture.
One type of software architecture used at the beginnings of MMD is the
blackboard: the knowledge is brought together in a kind of database that is
accessible to all the other modules, a bit like we have just illustrated for
real-time analyses. Each module can be dependent on restrictions covering
components in the database, and may only see a part of them. In that case, the
software architecture has a supervisor that determines the knowledge to
activate at every moment. Compared to chain implementation, blackboard
provides more flexibility and allows various modules to collaborate.
Obviously, it is necessary for the processes generated by the modules to allow
this type of collaboration: if the modules are implemented in the same way
as the chain model, these processes can only happen in a chain, whether the
data they require are accessible in the blackboard or whether it travels from