Zhao Han and Tom Williams


Companion Proceedings of the 17th ACM/IEEE International Conference on Human-Robot Interaction (HRI LBRs)

Publication Year


Best Late Breaking Report Nominee (3rd Place)

In many domains, robots must be able to communicate to humans through natural language. One of the core capabilities needed for task-based natural language communication is the ability to refer to objects, people, and locations. Existing work on robot referring expression generation has focused nearly exclusively on generation of definite descriptions to visible objects. But humans use many other linguistic forms to refer (e.g., pronouns) and commonly refer to objects that cannot be seen at time of reference. Critically, existing corpora used for modeling robot referring expression generation are insufficient for modeling this wider array of referring phenomena. To address this research gap, we present a novel interaction task in which an instructor teaches a learner in a series of construction tasks that require repeated reference to a mixture of present and non-present objects. We further explain how this task could be used in principled data collection efforts