Introduction
In human conversation, participants frequently use "fillers" such as "OK," "Got it," and "Mm-hmm." While these utterances might appear to be mere conversational noise or polite habits, Cohen and Levesque (1991) argue that they are a functional necessity for collaborative work. By applying a formal logic of shared commitment, the authors demonstrate that these confirmations are the "logical glue" that allows agents to successfully discharge or update joint goals.
The Problem: Beyond Individual Plans
During the late 1980s, AI discourse theories were primarily "plan-based." These models viewed dialogue as a process where one agent attempts to infer the hidden plan of another. However, these individualistic models struggled to explain the constant stream of confirmations observed in real-world tasks.
If an expert instructs an apprentice to "attach the pink tube," and the apprentice performs the action, why do they bother saying "Got it"? Under standard plan-based theories, the goal is achieved once the action is performed; there is no clear reason why the agent would then form a new goal to speak. This gap exists because these theories treat dialogue as two individuals acting in parallel rather than as a unified team engaged in a joint activity.
As shown in the dialogue above, participants often provide feedback even before a full instruction is completed. This suggests that the "goal to speak" is not just about reporting completion, but about maintaining a shared state.
A Logic of Joint Commitment
The authors address this by employing a formal modal logic framework (BDI logic—Belief, Desire, Intention) to define the mental states of agents. They introduce two critical concepts:
- Joint Persistent Goal (JPG): A state where a group mutually believes a goal is not yet achieved but remains worth pursuing. Crucially, if one member discovers the goal is finished (or impossible), they inherit a "residual goal" to make that discovery mutual knowledge.
- Joint Intention (JI): A JPG where the agents also mutually believe they are about to perform the action to achieve it.
In this framework, the "residual goal" is the trigger for confirmation. When a team has a joint goal and one person completes their part, the joint commitment does not simply vanish. Instead, the logic forces the person who finished to ensure their partner knows it is done. We say "Done" to release our partner from the joint commitment.
Key Insights: The Functional Necessity of Confirmation
The paper identifies two primary types of confirmations that emerge from this logic:
- Referent Identification: Confirming that the listener knows which object the speaker is talking about.
- Successful Action: Confirming that a physical step in the joint task has been completed.
Because joint action cannot proceed without mutual belief of success, confirmations are not just predicted; they are normatively expected. If a confirmation is missing, the initiator of the action will often explicitly request one (e.g., "You see that?" or "You got that on, okay?"), as seen in the figure above.
Empirical Validation
The authors analyzed task-oriented dialogues from a study involving experts and apprentices assembling a toy water pump. The data revealed that confirmations and requests for them accounted for 18% of the total verbal interaction time, occurring on average every 5.6 seconds.
Quantitatively, the model proved highly accurate: 89% of all confirmations in the telephone dialogues fell into the two categories predicted by the Joint Intention model (action success or referent identification). This provides strong evidence that dialogue is best modeled as teamwork rather than individual plan recognition.
Conclusion and Future Work
Cohen and Levesque’s work shifted the focus of discourse modeling from individual intentions to joint action. While their "possible worlds" logic faces challenges in scaling to general, non-task-oriented conversation, it provides a rigorous foundation for understanding why we talk the way we do when working together. Future research remains to be done on the computational complexity of these mental state updates and the application of this logic to casual, non-instrumental conversation.
Reference: Cohen, P. R., & Levesque, H. J. (1991). Confirmations and Joint Action. Proceedings of the 29th Annual Meeting of the Association for Computational Linguistics.
Summary updated