Intent Carryover

Intent carryover is fundamentally dependent on a robust Dialogue State Tracker. The DST maintains a structured representation of the user's goal across turns. When a user says 'What about tomorrow?', the DST references the slot-value pairs from the previous turn (e.g., intent: book_flight, destination: London) to resolve the anaphoric reference. Without DST, the system treats each utterance as a cold start, losing the established intent.

A primary function of intent carryover is enabling multi-turn slot filling. The system recognizes that a fragmentary utterance is not a new intent but a continuation of the parameter collection for the active intent.

• Turn 1: 'Book a flight to Tokyo' (Intent: BookFlight, Dest: Tokyo)
• Turn 2: 'On Friday' (Carryover: Updates date slot for the active BookFlight intent)
• Turn 3: 'Make it business class' (Carryover: Updates class slot)

This prevents the system from interpreting 'On Friday' as a standalone, unclassifiable statement.

Standard coreference resolution links pronouns to entities, but intent carryover extends this to link elliptical utterances to the prior speech act. The system must map 'How about the 5th?' not just to the date entity, but to the specific search_availability action previously invoked. This requires a discourse-level understanding that the user is proposing an alternative parameter for the same underlying task.

Intent carryover is constrained by the context window. The system prompt and the initial turns establishing the intent must remain within the model's active attention span. If a long conversation causes context window truncation that evicts the original intent declaration, the carryover mechanism fails. This necessitates strategies like contextual compression or summarizing the active goal in the system state to preserve it beyond the raw token limit.

Robust systems implement a decay function for carried-over intents. A user might abandon a booking flow and start a new task. The system must detect a topic shift or a new explicit intent that overrides the stale state.

• Explicit Override: 'Actually, cancel the flight and find me a hotel.'
• Implicit Expiry: A timeout threshold (e.g., 5 minutes of inactivity) resets the active intent.
• Conflict Detection: A new utterance with a high-confidence, distinct intent classification overrides the carried-over intent.

Intent carryover is a controlled, desired behavior that maintains goal orientation. In contrast, context drift is a degradation where the model gradually shifts focus unintentionally. Carryover keeps the system anchored to the original objective, while drift causes the system to lose sight of the primary goal. Effective carryover mechanisms actively resist drift by re-anchoring the model's attention to the active intent stored in the session state.

The systematic process of estimating the user's current goal and belief state at every conversational turn. DST aggregates the full dialogue history to maintain a probability distribution over possible intents and slot-value pairs. When a user says 'What about tomorrow?', DST identifies that 'tomorrow' updates the date slot of the previously captured booking intent without requiring the user to restate 'I want to book a flight.' Modern approaches use encoder-decoder architectures to jointly model intent classification and slot filling, replacing brittle rule-based systems.

The NLP task of identifying all linguistic expressions that refer to the same real-world entity. This is the underlying mechanism that enables intent carryover by linking pronouns and anaphora to their antecedents. Key resolution types include:

• Pronominal anaphora: resolving 'it', 'they', 'that one'
• Definite noun phrases: resolving 'the document', 'that policy'
• Zero anaphora: resolving omitted subjects common in pro-drop languages Without robust coreference resolution, a system cannot understand that 'Send it to legal' refers to the contract mentioned three turns earlier.

The task of extracting specific parameters required to execute a user's intent across multiple conversational turns. Each intent defines a schema of required and optional slots. Intent carryover preserves the active slot-filling frame so that subsequent utterances populate remaining slots without re-declaring the intent. For example:

• Turn 1: 'Book a flight to London' → Intent: BookFlight, Slot: destination=London
• Turn 2: 'Make it business class' → Slot: cabin_class=business
• Turn 3: 'Actually, Paris instead' → Slot: destination=Paris The system maintains the BookFlight frame across all turns, accumulating and revising slot values.

The gradual deviation of a conversation from its original topic or intent, representing a primary failure mode for intent carryover systems. Drift occurs when the model fails to distinguish between a legitimate intent modification and an unrelated topic shift. Mitigation strategies include:

• Intent confidence thresholds: decaying the active intent when slot-filling relevance drops below a calibrated score
• Explicit topic segmentation: detecting semantic boundaries where the conversation shifts to a new domain
• Clarification triggers: prompting the user when ambiguity exceeds a defined threshold Unchecked drift causes the system to apply new utterances to stale intents, producing nonsensical composite actions.

A critical failure state where the model loses the distinction between different conversational threads or temporal states, flattening the dialogue into a single incoherent prompt. In the context of intent carryover, collapse manifests when the system conflates a resolved intent with the current active intent. For example, after booking a flight and starting a hotel search, a collapsed context might apply 'near the airport' to the completed flight booking rather than the active hotel search. Prevention requires temporal state segmentation and explicit intent lifecycle management with clear activation and resolution markers.

The ability to fork a dialogue from a specific prior turn to explore alternative paths without corrupting the original session state. This is essential for intent carryover in complex scenarios where users say 'Actually, go back—what if I want a refund instead?' Branching creates an immutable dialogue tree where each node preserves the exact context and active intents at that point. The system can traverse back to a prior node and resume with a modified intent while retaining the ability to return to the original branch. Implementation typically uses event-sourced session stores rather than simple linear histories.