Exploring Strategies for Conceptual Alignment in LLM-Based Human-Robot Dialogue

Shengchen Zhang, Meiying Li, (Melody) Zixuan Wang, Xiaohua Sun, Weiwei Guo

Abstract

Successful conversations require speakers to align on conceptual understanding, a challenging but crucial task in human-robot interaction. With increasing use of large language models for dialogue, robots move from passively acquiring human conceptualizations to actively shaping alignment. However, the design space of such alignment dialogues and how different strategies are perceived and interpreted by people remains poorly understood. This paper presents a structured exploration of alignment strategies. First, we collect and analyze human dialogues from an alignment task to derive two design dimensions and four representative strategies. We then implement the strategies in a user study with a simulated robot, in a household organization scenario. Building on the findings, we develop the dimensions and strategies into a design space, with considerations for potential factors and trade-offs. We conclude with discussing how this design space can aid in the generation and analysis of alignment strategies, and implications for future research.

Cite as

TBA