Co-adaptation in situated dialogue reference resolution

Referring to objects in a common environment comprises a large part of human communication. However, while individual speakers’ usage of language for referring to entities is highly idiosyncratic, participants in dialogue show strong alignment in their use of referring language with that of the other participants: This means that, over time, the variance of referring language within a dialogue reduces despite that the variance between dialogues remains large. Moreover, most work on reference resolution in situated dialogue considers referring language as isolated, atomic referring expressions rather than as an iterative process of reference as it is observed in natural dialogue. This talk describes methods for modelling these phenomena in order to adapt referring-language semantics to individual dialogues and to automatically infer the ability of any given word to refer to objects in general, independent of the precise semantics of the word and thus circumventing the need to annotate referring expressions.