Semantic Analyses and Representations - Man-Machine Dialogue: Design and Challenges - page 84

Thus there are a multitude of ways to implement a syntactic analysis, just
like there are a multitude of representation formalisms. Depending on the
user's utterance, an implementation can turn out to be performing more than
another at certain moments and for certain phenomena. If we have the
required computational means, we can consider, as [VIL 05] did,
implementing several algorithms, i.e. running a “syntactic multi-analysis”.
The idea is to extract a trustworthy analysis from the results of various
analyzers, each result can be accompanied by a satisfaction score, just like in
the sketchboard approach seen in section 4.1.2. The observation behind this
approach is linked to the success of processes which combine multiple
outputs, both at the level of combining hypotheses made by the speech
recognition module or combining morphosyntactic labeling hypotheses
[VIL 05, p. 20]. If satisfaction scores cannot be calculated, one strategy is to
favor the results that produced the greatest number of times by all the
analyzers involved. In that case, it is useful to maximize the number of
analyzers, knowing that the real-time MMD constraints might slow down this
approach: even if the syntactic analysis is not the most greedy in terms of
memory and calculation time, it is good to be prepared.
5.2.2. Semantic and conceptual resources
The list of words in a language with their category and morphological
properties is the main resource necessary for syntactic analysis. Other
resources can also be used, such as statistical data on the word succession and
possible sentence structures. To go further and discuss the meaning,
additional data are required. It is first about the meaning of words, or at least
of a formal representation, for example a feature structure that puts aspects
into boxes, such as object types (concrete inanimate, abstract, animate and
human individuals), types of properties (gradable or non-gradable), events
(actions and processes, which are limited in time) and states (not limited in
time). The language dictionaries, unfortunately, cannot be used to fill such
structures for they are not structured enough to express the meanings in
natural language, which makes it a circular problem: we have to
automatically interpret the definitions of a dictionary to obtain useful
resources for automatic interpretation.
Various approaches are used, if possible in a complementary way, to build
what is called a semantic lexicon. The component approach aims to specify,