Reference Resolution - Man-Machine Dialogue: Design and Challenges - page 100

conceptual model, which turns out to be highly compatible with the fact that
it lasts a certain amount of time, which is the aim of the utterance's main
question. However, the referring expression “this itinerary” is not fully
analyzed. There is notably no referent for it yet. To achieve this, the
multimodal reference domain model determines an underspecified that
translates the linguistics constraints carried by the referring expression. These
constraints are first those of the words used, that is of the category and
modifiers as filters to search for the referent among the available objects. In
some cases, as in “color the pyramid red” or “delete this file”, the verb's
semantics and the sentence's semantics provide additional filters: the fact that
it is not red for “the pyramid” and the fact that it can be deleted for “this file”.
Another constraint is that of the scope of determination. Depending on the
demonstrative's way of functioning, the definite and undefinite, the referent
research criteria are different. Thus, the demonstrative “this N” forces the
focalization of the referent, either by previously mentioning the same referent
or simultaneously through a pointing gesture. As for the definite “the N”, it
operates by extracting the only N in the reference domain. As Corblin
[COR 95, p. 51] states, “the N” consists always of opposing a previously
mentioned N to other entities to predicate something about it. It has to oppose
the element that is an N in the reference domain to elements that are not N.
Finally, the undefinite operates by selecting a random element from a set.
These three cases are far from covering all the referring expression allowed
by language (plurals have their own mechanisms as do personal pronouns and
proper nouns), but they illustrate the three main mechanisms involved in
reference resolution: using focalization, extraction, selection.
At this point, the MMD system thus has a underspecified reference
domain with linguistic constraints on the referring expression, and this
underspecified domain will try and work with the reference domains provided
by the visual context, by analyzing the gesture if there is one, and by the
dialogue history. One of the roles it has is to save the successive reference
domains, so as to chart the phenomena or contextual broadening, contextual
narrowing, anaphora, as well as alterity, with expression of “other” such as
“the other trains” or “the other pyramid”. Depending on the task, the
matching can be tested in a specific order, by favoring, for example, visual
perception over the dialogue history and stopping as soon as a complete result
is achieved, or it can consist of clarifying all the possibilities, so that the
dialogue manager can decide which alternative to choose if there is any