Improving Grounded Language Acquisition Efficiency Using Interactive Labeling

Abstract

Natural language has emerged as a powerful, intuitive interface for robot-human communication. There has been substantial work in recent years on grounded language acquisition, in which paired language and sensor data are used to create a model of how linguistic constructs apply to the perceivable world. While powerful, this approach suffers from the need for extensive natural language annotations. In this paper, we describe an initial pilot of a system that uses active learning to solicit annotations from a human interlocutor. Our results suggest that using active learning reduces the number of annotations necessary to learn such groundings, providing a strong justification for building a more robust version of such a system, and suggest some insights into human requirements for usability.