State tracking

State Tracking (ST) is the core computational process that maintains a real-time, structured representation of a user’s goals, constraints, and requests throughout an interaction.

This technology, often called Dialogue State Tracking (DST) in AI, is mission-critical for task-oriented systems: it estimates the user’s belief state at every turn, typically as a set of slot-value pairs (e.g., 'destination: Paris', 'date: tomorrow'). The DST unit processes natural language understanding (NLU) outputs, updates the belief state, and feeds this structured data to the Dialogue Policy module for the next action. Modern approaches leverage deep learning models like BERT-DST and are validated on benchmarks such as the MultiWOZ dataset, with performance measured by Joint Goal Accuracy (JGA) — a metric that demands perfect state prediction per turn.

https://github.com/yukyunglee/Awesome-Dialogue-State-Tracking