Coherence, Symbol Grounding and Interactive Task Learning

To teach agents through natural language interaction, we need methods for updating the agent’s knowledge, given a teacher’s feedback. But natural language is ambiguous at many levels and so a major challenge is for the agent to disambiguate the intended message, given the signal and the context in which it’s uttered. In this paper we look at how coherence relations can be used to help disambiguate the teachers’ feedback and so contribute to the agent’s reasoning about how to solve their domain-level task. We conduct experiments where the agent must learn to build towers that comply with a set of rules, which the agent starts out ignorant of. It is also unaware of the concepts used to express the rules. We extend a model for learning these tasks which is based on coherence and show experimentally that our extensions can improve how fast the agent learns.