A Situated Context Model for Resolution and Generation of Referring Expressions

Abstract

The background for this paper is the aim to build robotic assistants that can "naturally" interact with humans. One prerequisite for this is that the robot can correctly identify objects or places a user refers to, and produce comprehensible references itself. As robots typically act in environments that are larger than what is immediately perceivable, the problem arises how to identify the appropriate context, against which to resolve or produce a referring expression (RE). Existing algorithms for generating REs generally bypass this problem by assuming a given context. In this paper, we explicitly address this problem, proposing a method for context determination in large-scale space. We show how it can be applied both for resolving and producing REs.