Intent Disambiguation

The system calculates a probability distribution over potential intents. If the top intent's score falls below a predefined confidence threshold (e.g., < 0.7), the agent triggers a clarification prompt rather than risking a wrong action.

• Top-1 vs. Top-2 Margin: Measures the gap between the highest and second-highest scoring intents
• Entropy-Based Gating: Uses Shannon entropy across the intent distribution to quantify uncertainty
• Dynamic Thresholds: Adjusts sensitivity based on the cost of error for the specific domain

When ambiguity is detected, the system must formulate a precise, non-leading question that efficiently narrows the intent space. Effective clarification questions present mutually exclusive options derived from the top-N competing intents.

• Slot-Filling Prompts: Asks for a missing parameter that disambiguates the goal
• Multiple-Choice Disambiguation: Presents 2-3 distinct intent candidates as selectable options
• Example: 'Did you mean account balance or recent transactions?'

Disambiguation extends across dialogue turns. The Dialogue State Tracker (DST) must retain the original ambiguous query, the clarification question issued, and the user's disambiguating response to update the belief state.

• Slot Carryover: Preserves previously extracted entities while resolving the ambiguous slot
• Intent Revision: Overwrites the original ambiguous intent with the clarified selection
• Turn-Level Annotation: Logs the disambiguation event for conversation analytics and model fine-tuning

Ambiguity is often resolved without explicit clarification by leveraging contextual signals from the user's session, profile, or environment.

• Session History: Prior turns constrain the intent space (e.g., a follow-up about 'it')
• User Profile: Role-based defaults (e.g., an admin vs. a viewer) pre-resolve ambiguous commands
• Geotemporal Context: Location and time of day disambiguate queries like 'What's open near me?'

Ambiguity often stems from polysemous entities—terms that refer to multiple distinct real-world objects. The system links the ambiguous mention to a knowledge graph to enumerate candidate entities.

• Candidate Generation: Retrieves all entities matching the surface form (e.g., 'Paris' → city, person, film)
• Attribute Comparison: Uses entity properties (type, popularity, context) to rank candidates
• Clarification via Attribute: Asks 'Did you mean Paris, France or Paris, Texas?' using distinguishing attributes

When a disambiguation guess is wrong, the system must capture the correction signal to update its intent classification model and prevent recurrence.

• Implicit Feedback: User abandons the session or reformulates the query after a wrong action
• Explicit Correction: User states 'No, I meant X' — this is a high-value training signal
• Active Learning Loop: Ambiguous cases with corrections are flagged for human annotation and model retraining