jasonwu0731/trade-dst

Built with PyTorch, TRADE uses a shared encoder-decoder architecture with a copy mechanism and slot gate to generate dialogue states across multiple domains, enabling zero-shot and few-shot transfer to unseen domains. The model decodes (domain, slot, value) triplets independently while leveraging a gating mechanism to predict slot activation, addressing the key limitation of tracking unknown slot values not seen during training. Includes support for continual learning approaches (EWC, GEM) to adapt to new domains without catastrophic forgetting, evaluated on the MultiWOZ dataset.

394 stars. No commits in the last 6 months.