Semantic Analyses and Representations - Man-Machine Dialogue: Design and Challenges - page 76

The confusion between “seems” and “seams”, which can happen in an oral
dialogue shows us an example of two homophones. Homophones are words
that are pronounced in the same way, whereas they have different meanings.
Furthermore, language has the particularity that a single word can have
various meanings, a phenomenon called polysemy. Thus, all the words of
language, or almost all of them, are polysemic. The train ticket that we have
mentioned earlier can refer to the piece of paper or the right to a seat which it
represents, and if we just limit ourselves to the word “ticket”, without taking
into account the compound noun “train ticket” the number of possible
meanings has gone beyond 10. Polysemy comes from metonymical or
metaphorical relations (see later), from meaning restrictions, extensions
(“minute”, refers to a very specific amount of time, but also, by extension, to
a very short period of time) or even phenomena linked to the differentiation of
a feature that leads us to take two notions into account. For NLP and open
domain MMD, there is a challenge in describing these transformations (or
using a parallel path based on the statistics on joint occurrences) so as to
automatically deduce the possible polysemic meanings of a starting point. For
a closed domain MMD, we will have to suffice with multiplying the lexicon
entries and keeping the most relevant meanings when taking the task into
account.
Language also has the particularity that one word can take the place of
another word. Metonymy thus takes place when a container takes the place of
thecontent(“trainreservation”foratrainseat),aparttakestheplaceofawhole
(synecdoche), a quality takes the place of a person (“first class is very noisy”),
an instrument takes the place of an agent, an action takes the place of an agent,
of the place it is carried out in, of its own effect, etc. For MMD, we will find
the two previous techniques: either the lexicon entries and potential sentence
structures are multiplied, which leads to a multiplication of possibilities which
is manageable when it comes to a specific task, or adding rules allowing for
metonymy replacements which enable a fine semantic analysis.
Metaphor is a close phenomenon that consists of giving one word
another's meaning, by analogy. Thus the buttons and menus of MMI are
called metaphors. If we keep with the train ticket reservation dialogue, the
following sentence, “I don't want to ride a snail, I will take the train going
through Meudon”, is a hard sentence for the system to interpret if, due to the
nature of the task, “snail” is not a word in its lexicon. In that case, one of the
challenges for NLP or open domain MMD is to find the feature of the “snail”
that will allow the system to understand its relation to a train, and thus