Reference Resolution - Man-Machine Dialogue: Design and Challenges - page 92

Chapter 6
Reference Resolution
Reference resolution as well as the processes that we will study in
Chapters 7 and 8 require input arguments, which can be preprocessed, and
give an output result, which might require postprocessing. We are here in the
presence of input information, after a pre-interpretation process. We are then
in the presence of one or more written transcriptions of the oral utterance
(several if there is any ambiguity), with prosodic indications, and simplified
representations for gestures. For multimodality to be correctly managed, these
transcriptions and representations are accompanied with temporal markers,
for example the date stamp of the beginning and end of each word
pronounced, of each prosodic accentuation and each gesture trajectory. These
aspects are a first set of arguments for reference resolution. A second set
brings together the results of lexical, syntactic and semantic analyses as they
have been presented in Chapter 5. Finally, a third set of arguments is the
dialogue history, in the cases when a linguistic expression refers to a previous
reference, which is the case for pronoun and associative metaphors. As is the
case for the utterance being processed, the history has two types of
representations: representations stemming from linguistic analyses and
representations as chain lists of words with temporal and prosodic markers.
The first type of representations are crucial to interpret “it” in “put it in the
box” when it follows “take a green block”, and the second type of
representations are to interpret “what I thought was a pyramid” or “what I
called a strange shape”, i.e. for the cases where the referring expression
contains properties that do not belong to the referent and cannot be accessed
through it.