Conversational Grounding as Natural Language Supervision -- the need for divergent agent data

Oliver Lemon

TL;DR: Argues for exploration of `conversational grounding' supervision signals in learning, and new data collections/tasks to enable this.

Abstract: We explore how Conversational Grounding messages in Natural Language can provide general and detailed feedback mechanisms for learning. We first present the fine-grained and targeted feedback signals provided by Conversational Grounding and discuss their potential advantages in models of language and task learning. We argue that a key factor holding back research in this area is lack of appropriate data on tasks with divergent agents, which can resolve disagreements and errors, and we propose requirements and methods for new data collections enabling such work.

Track: Non-Archival (will not appear in proceedings)