BossNet

BossNet: Disentangling Language and Knowledge in Task Oriented Dialogs

Datasets

The complete data folder structure with all tasks can be downloaded here.

Individual Datasets

• bAbI Dialog (26.2 MB)
• CamRest (1.1 MB)
• Stanford Multi Domain Dataset / SMD (6.7 MB)

Run Environment

We include a requirements.txt which has all the libraries installed for the correct run of the BossNet code. For best practices create a virtual / conda environment and install all dependencies via:

❱❱❱ pip install -r requirements.txt

Training

The model is run using the script main.py

❱❱❱ python main.py --task_id 1 --train 

The list of parameters to run the script is:

• --task_id this is task dependent. 1-5 for bAbI, 7 for CamRest, and 8 for SMD
• --batch_size batch size
• --learning_rate learning rate
• --embedding_size hidden state size of the two rnn
• --hops number of stacked rnn layers for BossNet
• --word_drop_prob dropout rate
• --p_gen_loss_weight loss function weight on copy

Look at params.py for detailed information on the runtime options

Training from Saved Model

There is support to start training from a previously saved checkpoint with the --save flag.

Testing

To obtain metric scores on the best model run main.py with --train=False or by omitting the train flag altogether. Make sure all the parameter options match those of the trained model.

❱❱❱ python main.py --task_id 1

Analysis

Trainable Paramters

The following numbers are reported using embedding_size 256, batch_size 64, hops 6

Running Times

The following numbers are reported using embedding_size 128, batch_size 64, hops 3
Times are reported as sec. per train epoch (avg. no. of epochs till convergence) **total train time**

Hyperparameters