[Requirements Engineering] ITM-SN: Solving the Multi-User Negotiation Puzzle for Conversational AI

TL;DR

As intelligent assistants move from simple personal tasks to complex group coordination (like planning a group trip), they face a "social needs" challenge. This paper proposes the Intention Tree Model for Social Needs (ITM-SN), an extension of the personal intention tree. It introduces a mechanism to track consensus across multiple users, manage key vs. optional participants, and drive the negotiation process from initial individual desires to a final, unified group agreement.

Background & Motivation: Beyond "My" Needs to "Our" Needs

Traditional requirement modeling is inherently "self-centered." Whether it's a simple list or a Personal Intention Tree (ITM-PN), the system assumes one user and one set of constraints.

However, real-world scenarios are social. Imagine four friends planning a trip to Beijing. User A wants a luxury hotel; User B needs to stay near the Great Wall; User C is on a budget. Traditional models break down here because:

1. They don't record who proposed which constraint.
2. They lack negotiation logic (e.g., "IsAgree" or "Union" operations).
3. They cannot identify consensus vs. ambiguity in real-time.

Methodology: The Architecture of Consensus

The core innovation of ITM-SN is the decoupling of the requirement element from the user identification, then re-linking them via a UserGroup.

1. The Meta-Model Expansion

In ITM-PN, an intention node was just a functional goal. In ITM-SN, every intention, constraint, and optimization objective is a tuple: $Intention=<intention,UserGroup>$ This allows the system to see that while "Visit Great Wall" might have 100% agreement, the constraint "Duration: 3 hours" might only have User A and B's support, while User C prefers "2 hours."

2. Model States & Transition

The authors define three critical delivery states for the model:

• Initial Model: Contains only the "origin user's" personal needs.
• Intermediate Model: A "living" tree where conflicting nodes (same name, different values) coexist, and users are added/removed from groups as they negotiate.
• Termination Model: The "Gold Standard" state where all Key Users have participated, no new needs are proposed, and no conflicting nodes remain.

3. Negotiation Strategies

• Serial Negotiation: The bot talks to Users A, then B, then C, refining the tree step-by-step.
• Parallel Negotiation: The bot facilitates a "round-robin" or group discussion where all users can see the current state and offer counter-proposals simultaneously.

Experiments: The "Beijing Tourism" Case Study

The authors validated ITM-SN using a parallel negotiation scenario involving four users (Original user, Tony, John, Amy).

The experiment tracked how the tree evolved:

• Round 2: Tony joins and adds a conflicting "Home Inn" preference.
• Round 3: John adds a "Bath" requirement but Tony disagrees.
• Disambiguation: The system identifies that Tony's refusal to agree on "Bath" prevents a Termination state. Through negotiation, the "Bath" constraint is deleted to achieve consensus, or a compromise is found.

(Visual representation of the ITM-SN during the second round of dialogue, showing the emergence of conflicting nodes in blue).

Critical Insight & Future Outlook

The primary value of ITM-SN is its mathematical rigor in defining social consensus. By distinguishing between Key Users (whose agreement is mandatory) and Optional Users, it provides a logical "exit condition" for conversational bots that previous models lacked.

Limitations: The paper currently focuses on the structure of the requirements. In the era of Large Language Models (LLMs), the next step will be mapping these structured "Intention Trees" to the messy, unstructured natural language of group chats.

Future Work: The authors aim to implement this into a functional Human-Machine dialogue system, likely leveraging the structured state transitions of ITM-SN to guide the LLM's dialogue policy.