Why do standard dialogue systems fail at tracking negotiation agreement?

Dialogue state tracking (DST) is the backbone of task-oriented dialogue — extracting user goals as slot-value pairs (e.g., "restaurant", "area", "centre"). But standard DST has a structural assumption: it tracks ONE user's goals. The system is a service provider filling slots for the customer.

Negotiation dialogue breaks this assumption. Agreement tracking requires monitoring BOTH interlocutors' commitments across multiple issues simultaneously. An employer and candidate negotiate salary, hours, and promotions — agreement on any issue requires explicit confirmation from both sides, not just one.

This is harder than standard DST for several reasons:

• Standard DST estimates goals of a single interlocutor; agreement tracking requires tracking two interlocutors' evolving positions
• Zero-shot and few-shot DST models, even those designed for unseen domains, are limited to form-filling paradigms (restaurant reservations, hotel bookings)
• The dialogue dynamics are fundamentally different: negotiation involves strategic information withholding, concession patterns, and bilateral commitment — not just information provision

The scarcity of annotated multi-issue negotiation corpora compounds the problem. GPT-NEGOCHAT uses GPT-3 to synthesize training data, but this introduces a dependency on synthetic data quality for a task where the interaction dynamics matter most.

Since Can AI systems detect when they've genuinely reached agreement?, agreement detection is valuable not just for negotiation but for any multi-agent deliberation. The bilateral requirement generalizes: whenever two or more parties must explicitly converge, tracking one side's state is insufficient.

Source: Conversation Architecture Structure