Section: New Results
Reference Resolution in Dialogue Systems
Reference resolution is a critical ability in a dialogue system because the natural behaviour of a speaker is to use short expressions designating the objects instead of full descriptions. The difficulty lies in the diversity of referring expressions (pronominals, definites, demonstratives, indefinites...) and the different modalities used to refer: gestural or linguistic. The Reference Domain model  , integrates the range of referring expressions in a multimodal framework, unlike the discourse representation theory, which focuses on a monomodal representation  .
However, the Reference Domain model had to be extended in order to manage plural referents and groupings. The problem was that the theory did not include a high level reference domain, acting as a container, for plural referents or newly created groupings. We added such upper-level domain and modelled the resolution of particular expressions like ordinals or alterities within the extended model. This work has been validated in the MEDIA understanding evaluation campaign, which gave a chance to test the model on a large spoken corpus (more than 1,200 dialogues). Indeed, the model was applied to several complex cases involving plural anaphoric or generic expressions which were hard to resolve, but proved to be efficient on the more usual expressions.