Transferable Dialogue State Generator (TRADE) model is a dialogue state tracker for multi-domain goal-oriented dialogue systems. This model is introduced by C. Wu et al. in https://arxiv.org/abs/1905.08743. This is a checkpoint for TRADE model trained on MultiWOZ 2.1 dataset using NeMo. It was trained for 10 epochs with batch size of 16 on single GPU with 16GB memory. The model achieves joint goal accuracy of 47.77% on the test set and 50.85% on the development set on MultiWOZ 2.1.
The files included in this model are:
This model is trained using NeMo. The source code and developer guide of NeMo is available at https://github.com/NVIDIA/NeMo. The example used to train the model is https://github.com/NVIDIA/NeMo/blob/master/examples/nlp/dialogue_state_tracking/dialogue_state_tracking_trade.py. You may found more information on the model and how to train it with this example in this toturial: https://nvidia.github.io/NeMo/nlp/dialogue_state_tracking_trade.html.
Usage example: Download the checkpoint files and place them in a checkpoint folder ($checkpoint_folder$). Download and process the dataset as explained in the toturial, and move it into $data_folder$. Then, run dialogue_state_tracking_trade.py example to evaluate the model on the test set:
python dialogue_state_tracking_trade.py --work_dir=$checkpoint_folder$ --data_dir=$data_folder$ --eval_file_prefix=test
To evaluate on the development set, you may use "--eval_file_prefix=dev".