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

	Task 1	Task 2	Task 3	Task 4	Task 5	CamRest	SMD
BossNet	1204993	1198849	1478913	1715457	1943809	1968129	2187265
Mem2Seq	776030	780127	3426789	4508397	5049201	5274536	6880560
Seq2Seq + Copy	6379859	6378321	6870481	7083494	7167315	7223452	7504137
Seq2Seq	6905172	6903634	7395794	7608807	7692628	7748765	8029450

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**

	Task 1	Task 2	Task 3	Task 4	Task 5	CamRest	SMD
BossNet	38.1 (15) 571.5	65.2 (10) 652.0	142.4 (25) 3560	16.9 (2) 33.8	231.3 (6) 1387.8	113.5 (6) 681	1252 (10) 12520
Mem2Seq	10 (100) 1000	32 (30) 960	51 (90) 4590	4 (10) 40	136 (60) 8160	22 (40) 880	81 (40) 3240

Hyperparameters

	Task 1	Task 2	Task 3	Task 4	Task 5	CamRest	SMD
Learning Rate	0.001	0.001	0.005	0.001	0.0005	0.0005	0.0005
Hops	1	1	3	1	3	6	3
Embedding Size	128	128	128	128	256	256	256
Disentangle Loss Weight	1.0	1.0	1.5	1.0	1.0	1.0	1.0
DLD	0.2	0.2	0.2	0.2	0.2	0.2	0.1