Abstract
Effective human-robot interaction requires robots to navigate the social nuances of Indirect Speech Acts (ISAs). This paper presents a hybrid cognitive architecture that combines fast, rule-based idiomatic processing with deep, plan-based inferential reasoning. This dual-path approach enables robots to generate socially appropriate requests and interpret indirect responses, including those manifested as physical actions. By incorporating social roles and politeness norms, the system ensures that robotic communication is both functional and socially acceptable.
Introduction
In natural human conversation, speakers often use Indirect Speech Acts (ISAs) rather than direct commands. For instance, "My hands are full" may function as a request to open a door. For a robot to be a competent social partner, it must master two primary challenges: generating requests that respect social hierarchies and politeness norms, and understanding indirect responses that may deviate from expected linguistic patterns.
Existing robotic systems typically rely on either rigid lookup tables (idiomatic) or computationally expensive reasoning (inferential). This work proposes a hybrid model that leverages the efficiency of idiomatic processing for conventional ISAs while utilizing plan-based reasoning for unconventional scenarios and non-linguistic responses.
The Hybrid Architecture
The proposed system is implemented within the Agent Development Environment (ADE). The architecture features two primary pathways for processing speech acts:
- The Idiomatic Path: Conventional ISAs (e.g., "Can you...?") are handled by a Pragmatic Reasoner. This component applies pre-defined rules to quickly map surface speech to intended meaning, bypassing the need for complex planning.
- The Inferential Path: For unconventional ISAs or indirect answers, the system employs a Planner/Plan-Reasoner based on the GraphPlan algorithm. This path treats communication as a sequence of actions designed to modify the listener's mental state.
This architecture allows the robot to handle "Case 3" events—unexpected actions that align with a valid alternative plan to achieve the robot's goal. For example, if a robot asks for directions and the human simply begins walking, the plan-reasoner identifies this physical movement as a valid step in a shared navigation plan.
Social Modeling and Politeness Norms
A key contribution of this research is the integration of social context into the reasoning process. Rather than assuming agents adopt goals simply because they are asked, the system models social obligations tied to specific roles (e.g., manager vs. co-worker).
The system utilizes a politeness metric (η) to evaluate the "face-threatening" nature of potential utterances. The robot selects the most direct utterance that does not exceed a socially-determined threshold for the given context.
As shown in the table above, the appropriateness of a request is highly dependent on the relationship and the prevailing norm. A direct request might be considered "non-politic" (too direct) for a manager when politeness is favored, whereas an unconventional indirect request might be perfectly appropriate for a co-worker.
Evaluation and Results
The architecture was evaluated through qualitative walkthroughs in simulated laboratory environments. The system successfully demonstrated:
- Appropriate Generation: The robot correctly adjusted its request style based on its social role. As a peer, it used "off-record" indirect requests; as a supervisor in a high-directness environment, it utilized direct commands.
- Robust Understanding: The robot successfully interpreted three types of responses to a location query: direct verbal answers, indirect verbal cues ("Follow me"), and purely physical actions (gesturing and walking away).
Conclusion
This hybrid approach provides a robust framework for socially-aware robotic communication. By balancing idiomatic efficiency with inferential flexibility and grounding both in social obligations, robots can interact with humans in a more natural and acceptable manner. Future work will focus on integrating advanced action recognition systems and conducting real-world human-robot interaction studies to validate these strategies.
Summary updated