An Empirically-based Model for Perspective Selection in Route-Finding Dialogues

In this work we aim to computationally model the extent to which certain empirical factors affect spatial perspective selection as used in route-finding dialogues. In such dialogues, both interlocutors need to adopt a spatial perspective in which to describe movement direction. In map-based tasks such as the one we are concerned with, two perspective choices are typically available, i.e., route perspective, where projective terms are defined with respect to the perspective of the route follower themselves, e.g., ``go to your right'', or survey perspective, where projective terms are defined with respect to a global or allocentric perspective, e.g., ``go down'', or ``go toward the top of the screen''. Addressees must be able to assign perspectives to a given spatial term in order to correctly interpret the utterance it is contained in. However the most frequent directional terms, i.e., `left' and `right', can be used in either route or survey perspective, and perspective is not typically marked explicitly at the lexical level. Generally addressees do correctly assign perspective to projective terms, even when perspective is not indicated explicitly in language, but misunderstandings may occur and clarification is often necessary.