An Explanation-oriented Inquiry Dialogue Game for Expert Collaborative Recommendations

Technical Summary: An Explanation-oriented Inquiry Dialogue Game for Expert Collaborative Recommendations

Problem Statement
The paper addresses the challenge of designing collaborative multiagent systems (MAS) where human and artificial experts must jointly generate recommendations, particularly in high-stakes domains like healthcare. Current literature often focuses on explaining system outputs to human users after a decision is made (post-hoc explainability). However, there is a lack of frameworks that incorporate explainability into the reasoning process of the agents themselves. The authors identify a need for a dialogue protocol that allows experts with different knowledge bases to collaboratively resolve conflicts, transfer knowledge, and generate a recommendation accompanied by a rich, traceable history of the reasoning process. Existing communication tools in healthcare (e.g., messaging platforms) often lack structured protocols to enforce the exchange of justifications and explanations, leading to potential communication failures and opaque decision-making.

Methodology
The authors propose a novel interaction protocol called the Experts' Dialogue Game (EDG). The methodology follows a user-centered software engineering approach:

1. Requirement Analysis: The authors conducted a systematic review of medical literature and informal discussions with medical experts to identify 14 requirements for effective expert collaboration. These were categorized into four classes:

   • Agent-Oriented: Mitigating individual communication constraints.
   • Cooperation-Oriented: Promoting open discussion, knowledge transfer, and quality control.
   • Protocol-Oriented: Ensuring explicit communication of history, critical points, explanations, and conflict resolution mechanisms.
   • Implementation-Oriented: Addressing coordination, recording, and privacy.

2. Formal Protocol Design (EDG):

   • Type: EDG is defined as an inquiry dialogue (where the goal is to find proof/evidence) enriched with explanatory illocutionary forces.
   • Locutions: The protocol defines four subsets of speech acts (locutions):
     • Informational: observation, verdict, advise, concern, assert.
     • Requests: wh-explain (requesting details), wh-justify (requesting backing for a stance), wh-clarify (requesting term definition).
     • Replies: explain, justify, clarify, agree, retract.
     • Management: prompt, end, pass.
   • Mechanism: Unlike standard inquiry dialogues that rely heavily on assert moves, EDG forces participants to use explanation-based locutions when challenging or supporting a claim. This generates "richer traces" of the reasoning process.
   • Commitment Stores: The system maintains individual commitment stores for each agent and a collective Agreement Store (AS). The AS only accumulates statements that have survived critical debate and received explicit agreement (or lack of challenge), ensuring the final recommendation is robust.
   • Politeness Rules: To prevent incoherence, the protocol enforces rules requiring that wh-requests be answered by the addressee before other moves are made, and uses prompt to manage outstanding issues.

3. Implementation and Evaluation:

   • A prototype web application was developed to enforce the EDG protocol, translating formal locutions into natural language interfaces (e.g., "Can you justify" for wh-justify).
   • A formative user study was conducted with six medical students acting as experts. They used the platform to collaboratively diagnose a patient with a thyroid disorder, with patient data distributed unevenly among them to test knowledge transfer.
   • The study evaluated the system against the initial requirements and gathered qualitative feedback on usability, effectiveness, and the utility of the explanation requests.

Key Contributions

1. Requirement Specification: The paper provides a domain-grounded but abstract requirement specification for expert collaboration, applicable beyond medicine to fields like aviation and engineering.
2. Explanation-Oriented Inquiry Dialogue: The authors introduce a dialogue game that embeds explanatory illocutionary forces (wh-explain, wh-justify, wh-clarify) directly into an inquiry dialogue. This shifts the focus from merely exchanging facts to exchanging the reasons behind facts, making the agent reasoning process itself explainable.
3. Methodological Integration: The work demonstrates a methodology of importing user-centered software engineering practices (requirement gathering, prototyping, formative evaluation) into the design of formal dialogue games.
4. Empirical Validation: The paper presents a user study confirming that the dialogue game meets the needs of medical experts for collaboration and that the structured exchange of explanations is perceived as valuable for building trust and clarity.

Results

• Requirement Satisfaction: The evaluation (Table 8) indicates that the EDG protocol and its implementation satisfy the majority of the identified requirements. Specifically, the protocol successfully enforces cooperation (RC1), allows for detailed discussion (RC3), and facilitates knowledge transfer (RC4).
• User Feedback:
  • Participants found the platform valuable for professional communication, noting that the structured requests for justification helped overcome social barriers (e.g., the fear of being rude when asking for clarification).
  • The classification of explanation requests was generally useful, though some participants found the term "justify" slightly aggressive compared to "explain."
  • The turn-taking mechanism was seen as both helpful for coordination and occasionally restrictive, with suggestions for more dynamic turn management.
  • The separation of dialogue topics (history, diagnosis, advice) was rated highly for efficiency.
• Trace Generation: The study confirmed that the protocol successfully generates a collective agreement store that reflects the consensus reached through explicit reasoning, rather than just a list of unchallenged assertions.

Significance and Claims
The paper claims that EDG represents a significant step toward Explainable Artificial Intelligence (XAI) in multiagent systems. By integrating explanation into the reasoning phase rather than just the output phase, the system produces a transparent audit trail of how a recommendation was derived. The authors modestly state that while the dialogue game is formally grounded, its true novelty lies in its empirical grounding in user requirements and its application to real-world expert collaboration scenarios.

The work does not claim to solve all issues of multiagent coordination (e.g., it notes that runtime termination guarantees and strategies for non-cooperative agents remain open problems). Instead, it positions EDG as a foundational protocol that can be extended to hybrid human-artificial systems, offering a structured way to handle the complexity of interdisciplinary decision-making while maintaining the integrity and explainability of the process. The authors conclude that such dialogue-based tools have real-life value for the medical community, potentially improving the quality and transparency of collaborative diagnoses.