An Assessment of the Evolution of Research and Systems - Man-Machine Dialogue: Design and Challenges - page 20

and following their syntactic relations (propositional semantics); the
allocation of referents to the first and second person pronouns, to referential
expressions in general, and to anaphora in particular (pragmatic analysis
sometimes called first-level pragmatics); the identification of the implicit and
the context attached to the utterance (second-level pragmatics); and the
determination of speech acts so the system can understand the nature of the
user's intervention (third-level pragmatics). Beyond the simple transcription
of the literal meaning of an utterance, here we enter into the field of the
determination of its contextual meaning. As for NLP, the implemented
methods and algorithms have evolved for all these analyses. Where at one
time, the symbolic approaches stemming from the AI were the only ones
around, we now see statistical approaches, and they are at all levels of the list
given above. These approaches have proved their efficiency in many fields of
NLP, and have sometimes completely replaced symbolic approaches. In the
MMD domain, it is the hybridization of symbolic and statistical approaches
that provides us with the most promising results.
Starting with a semantic representation that is faithful to the utterance, we
then reach an enriched representation through one or more implicit or explicit
messages that the utterance carries forth. This enriched representation is what
the system will confront, internally, to previously manipulated representation
as the dialogue advances. It is also due to the information it contains that the
system will be able to abstract the structure of the dialogue and compare it to
structuresthatareconsideredforthistask.Thisapproachwasimaginedbackin
the 1990s [REI 85] but its computer-based implementation within a real MMD
system framework only happened much later, and is still going on today. The
system can thus carry out an assessment of the task's satisfaction, identify the
deficiencies and decide what its next intervention will be. All this proceeds
from the reasoning that it implements so as to process the user's utterance in
the more relevant manner, taking into account what has already been done,
what the utterance brings to the dialogue, and what still needs to be achieved
to satisfy the task. Here we are situated not in linguistics and pragmatics but
in modern AI: the themes approached are those of knowledge representation
and especially following formalisms stemming from logic, in order to allow
automatic deductions and those of expert systems and multicriteria decisions.
Actually, as for the analyses mentioned in the previous paragraph, we are faced
with a hybridization of various approaches. As an example, the approaches
based on the expressive power of a well-defined logic, and its consistency with