Abstract
In the development of collaborative artificial agents, a significant challenge lies in modeling the iterative process of negotiation. While existing frameworks like "SharedPlans" describe the state of mutual agreement, they often overlook the communicative dynamics required to reach that state. This paper introduces a formal artificial discourse language designed to bridge this gap. By utilizing specific message primitives and discourse stacks to track the evolution of beliefs, the proposed model provides a computational framework for agents to propose, reject, and harmonize intentions within a collaborative task.
Introduction
In the early 1990s, as researchers formalized how artificial agents might cooperate, they encountered a persistent hurdle: the "black box" of negotiation. While models existed to describe the end state of a shared plan, the messy, iterative process of talking one's way into an agreement remained poorly understood. In her 1994 AAAI paper, Candace L. Sidner introduced a formal, artificial discourse language to provide a roadmap for how agents can synchronize their beliefs and intentions to achieve common goals.
Collaboration requires more than just shared objectives; it demands a constant alignment of internal states. In human conversation, this alignment occurs through negotiation—a series of exchanges where ideas are tested, refined, or discarded. Sidner identifies a gap in the prevailing "SharedPlan" models of the era: while they described what agreement looks like, they failed to explain the communicative process of reaching it. To build truly collaborative agents, the complexity of human dialogue must be abstracted into a logical set of primitives that an automated system can track and manipulate.
The Artificial Language of Negotiation
Sidner’s approach defines an Artificial Language of Negotiation that moves away from processing raw natural language in favor of formal message types. Each message type has a defined effect on the "state of communication," allowing agents to track the progress of a negotiation mathematically.
Message Primitives
The language is built upon several core primitives:
- PFA (ProposeForAccept): Initiates a proposal for a new belief or action.
- RJ (Reject): Explicitly turns down a proposal.
- AP (AcceptProposal): Signals formal agreement with a proposal.
- CO (Counter): Provides an alternative to a previously discussed proposal.
- AR (AcknowledgeReceipt): Confirms that a message was heard without necessarily agreeing to its content.
Discourse Stacks and State Management
To track the "attentional state" of a conversation, the model employs two primary stacks:
- Open Beliefs: This stack contains proposals currently under consideration.
- Rejected Beliefs: This stack tracks proposals that have been explicitly dismissed, preventing the agents from looping back to failed ideas without new information.
When an agent sends an AP (Accept) message, the proposal moves from the "Open" stack into the set of Mutual Beliefs (MB). This transition is managed by an underlying belief revision engine, ensuring that both agents remain "in sync" regarding their shared commitments.
Key Insights in Discourse Logic
A pivotal contribution of Sidner’s work is the formal distinction between Acknowledgment and Acceptance. In human dialogue, a word like "OK" often serves merely as an acknowledgment (AR)—a signal that the listener heard the speaker. It does not always imply acceptance (AP). By separating these functions, Sidner’s language prevents agents from prematurely assuming a proposal has become a mutual belief, a common failure mode in simpler communication models.
Furthermore, the language accounts for complex discourse moves such as the "Call for the Question." This occurs when an agent asks a clarifying question to delay accepting a proposal until more information is available. This behavior is formalized as a method for managing the "Open" stack, allowing multiple proposals to remain active simultaneously as agents weigh different solutions.
Evaluation and Results
As a theoretical framework, the paper validates its model through two primary methods:
- Human Dialogue Transcription: The author analyzed real-world dialogues (specifically a planning session between two individuals, M and K) to ensure the artificial language could logically reconstruct the essential negotiation behaviors found in human speech.
- Simulated Dialogues: Sidner constructed synthetic dialogues between two agents (R and C) to demonstrate how the language handles intricate scenarios, such as counter-proposals and the retraction of previous statements.
The results demonstrated that a small set of formal primitives could effectively capture the logic of complex negotiation. The use of stacks provided a clear computational path for an agent to determine its next move based on the current state of the discourse.
Limitations and Future Directions
While foundational, the model operates under the "No Deception" caveat, assuming agents are strictly collaborative and honest. It is not intended for adversarial negotiation, such as a commercial transaction, where agents might strategically hide their true beliefs. Additionally, the language does not yet account for "repairs"—the linguistic mechanisms used when a participant misspeaks or fails to understand a previous turn.
Future work suggested by Sidner includes testing the language's completeness across a wider variety of human conversations and exploring its implementation as a direct interface between humans and software agents. By formalizing the "art of the deal," this research laid the groundwork for more sophisticated, communicative AI capable of true partnership.
Summary updated