A Taxonomy of Conceptual Alignment in Human–Robot Dialogue

Conceptual alignment—the process by which speakers negotiate the meaning of concepts to achieve mutual understanding—is essential for effective human-robot interaction (HRI). However, current HRI research often treats this as a passive, unidirectional process where a robot learns from a human. This paper argues that such a view is insufficient and proposes a design-oriented framework that treats alignment as a dynamic, co-constructive dialogue.

The authors introduce a structured taxonomy to help researchers analyze and design for conceptual alignment. The taxonomy characterizes alignment dialogues along two primary dimensions:

• Triggers for Initiation: The authors identify three distinct triggers: Terminology (responding to vague or ambiguous words), Behaviors (addressing actions that conflict with established common ground), and Situational (proactive efforts to summarize or confirm consensus).
• Levels of Alignment: The framework categorizes the depth of conceptual negotiation into five levels: Evaluation and judgment, Framing and hierarchy, Definition and rules, Examples and cases, and Perception and interpretation.

To operationalize this, the authors provide a dialogue act schema—a set of interactional moves (e.g., 'Propose idea', 'Demand justification', 'State uncertainty')—that allows researchers to map how alignment is achieved in practice.

The framework was developed by synthesizing existing literature on grounding and negotiation with an empirical analysis of 48 human-human dialogues. By using human interaction as a baseline, the authors aim to capture a broader range of strategies than those currently implemented in robotic systems. This work provides a foundation for HRI designers to move beyond simple question-answering loops and toward more sophisticated, bidirectional communication strategies that account for the subjective and situational nature of human concepts.