Information Technology Reference
In-Depth Information
module to module. This is another difference between conceptual architecture
and software architecture: the blackboard is useful so long as the conceptual
architecture has planned for modules that share the same data. The GUS
system agenda is an example. In modern implementation, we should note that
the current blackboard, at least for a connected MMD system, can be ... the
web (especially with the success of cloud computing).
Sabah [SAB 97] suggested an interesting software architecture in the
mid-1990s, called a sketchboard. In this implementation, each module is able
to assess what it generates while taking into account the inputs it has received.
Thus, when the syntactic analysis module receives the transcription of an
utterance, it runs a syntactic analysis, generates a sketch and calculates a
satisfaction score in relation to this sketch. This satisfaction score is
transmitted to the other modules, more specifically to the module that
generated the data used in input. Depending on the score, it may restart its
work and generate a new transcription, transmit it and thus encourage the
syntactic module to generate a new, hopefully better, sketch. This is in fact an
extension of the blackboard with feedback. For the modules not to generate
the same analyses every time, noise is introduced in each stage. Other
methods can also be considered to replace this noise that may lack relevance:
taking out a setting used by the user or modifying the importance of a setting,
for example, such as prosody.
The most successful software architecture is probably the multi-agent
system. Each module is matched with an agent in charge of its interactions
with the rest of the architecture. Thus, this agent manages all the inputs and
outputs, and can also carry out similar processes to those in the sketchboard.
With such a materialization, any problem is solved by the converging
interaction of different agents. The TRIPS system, a successor of the
TRAINS system [ALL 95], is implemented as a multi-agent system, with
approximately the following agents: interpretation, dialogue management,
generation, discourse context, referential context, task, which all
communicate with each other. The exchange possibilities are thus very
numerous, and it is the interaction between agents that allows the dynamic
data to stabilize itself and create a satisfying result.
However, the need to manage the dialogue along various levels or priority,
especially - if we follow our example with the utterance resource - at a
real-time level very close to the speech turn management and a more
Search WWH ::




Custom Search