dialogue state tracking
**Dialogue state tracking (DST)** is the task of maintaining a structured representation of the **current state of a conversation** — tracking what the user wants, what information has been provided, and what remains to be resolved. It is a core component of **task-oriented dialogue systems** like virtual assistants, booking systems, and customer service bots.
**What the Dialogue State Contains**
- **Slots and Values**: Key-value pairs representing the user's requirements. For example, in a restaurant booking: `{cuisine: "Italian", party_size: 4, time: "7pm", location: null}`. Unfilled slots indicate information still needed.
- **User Intent**: The user's overall goal — booking, information query, complaint, modification, etc.
- **Dialogue Acts**: The type of each utterance — inform, request, confirm, deny, etc.
- **Conversation History**: Accumulated context from all previous turns.
**Why DST Is Challenging**
- **Coreference**: "Make it for 6 instead" — the tracker must understand "it" refers to the booking and "6" updates party_size.
- **Implicit Updates**: "Actually, let's do Thai" implicitly updates cuisine and may invalidate the previously selected restaurant.
- **Multi-Domain**: Conversations may span multiple domains — booking a flight, then a hotel, then a car — each with its own slot schema.
- **Error Propagation**: ASR (speech recognition) errors and NLU misunderstandings compound across turns.
**Modern Approaches**
- **LLM-Based DST**: Use large language models to extract and update dialogue state from conversation history — achieving state-of-the-art results with in-context learning.
- **Schema-Guided DST**: Define slot schemas declaratively and train models to generalize to new domains and slots not seen during training.
- **Hybrid Systems**: Combine rule-based tracking for simple slots with neural models for complex, context-dependent state updates.
DST is essential for building dialogue systems that can maintain **coherent, multi-turn conversations** and reliably track user needs across complex interactions.