Semantic Analyses and Representations - Man-Machine Dialogue: Design and Challenges - page 83

5.2.1. Syntactic analyses
Syntactic analysis consists of highlighting and representing the structure
of a sentence in computational data, with its constituents - the verb, the
subject, the direct object complement, etc. To this end, it requires a list of
language words with, in each entry, the category (verb and noun) and
morphological properties (gender, number and person). The latter helps it
manage the morphology at the same time as the syntactic analysis is running.
Thanks to the work of R. Montague [MUS 96], automatic understanding,
especially for written dialogue, took the path that consists of carrying out a
syntactic analysis of the sentence before a semantic analysis, which will itself
lead to the determination of the logical form or forms describing the possible
meanings. A pragmatic analysis that takes into account the contextual aspects
then enriches this semantic representation to achieve a propositional form, the
result of all the automatic understanding processes.
To process spontaneous oral dialogue, this process, which goes through a
global syntactic analysis of the sentence, is not always adapted: as we have
seen earlier, a dialogue utterance can be an incomplete sentence, and other
mechanisms have to be implemented for the system not to crash due to a
syntactic analysis which is impossible to run. We then have to implement
syntactic analyzers able to create partial analyses, i.e. able to manage the
underspecification of an actant, for example, or local analyses, which
generate a result with the available data even if it is incomplete.
Whether it is a global or a local or partial analysis, or even a macrosyntactic
analysis, they all supplement each other so as to use the input data as best they
can. The local analysis principle notably allows us to temper the importance
of syntax in the understanding process and to highlight the semantic analysis,
which, in a way, leads to the operations and requires local syntactic analyses
when necessary. This type of mechanism is more adapted to the characteristics
of oral language, such as distortion or fragmentation phenomena, or simply
the frequent presence of questions, a type of sentence that does not appear
very often in written texts, and is thus not well taken into account by analyzers
who are designed and trained on written texts. It can also turn out to be more
robust to potential errors in speech recognition. This is especially the case in
closed domain MMD: for a train reservation task, we almost know in advance
all the elements that a user query can cover, and these elements can thus help
the analysis according to a top-down approach and supplement the bottom-up
approach started with speech recognition.