liuchang97/multi-domain-belief-tracking

The implementation of the model proposed in the Large-Scale Multi-Domain Belief Tracking with Knowledge Sharing paper

Large-Scale Multi-Domain Belief Tracking with Knowledge Sharing

The implementation of the model proposed in the Large-Scale Multi-Domain Belief Tracking with Knowledge Sharing paper

Requirements

Python 3+ with pip

Setup

To install the python modules required to train and run the model

pip install -r requirements.txt

Preprocessing

To download and pre-process the multi-domain belief tracking dialogues and pre-trained word embeddings

python preprocess.py

Training

To train the model

python main.py train [--args=value]

Some of these args include:

• --num_hid: the size of the hidden layers - default is 50 (the paper uses 100)
• --bidir/--no-bidir: bidirectional vs forward only - default is bidirectional (only valid for a RNNs)
• --net_type: the type of the feature encoders, gru for GRU, cnn for CNN, lstm for LSTM - default is lstm (See the paper)
• --batch_size: the batch size - default is 64
• --dev: device use to train the model (cpu or gpu) - default is gpu
• --model_url: path to save the model or to resume training - default is models/model-1
• --graph_url: path to save the tensorboard graph for evaluation metrics (cross-entropy, accuracy ..etc) - default is graphs/graph-1

For example to train the cnn variant to the model with 8 batch size on the GPU

python main.py train --batch_size=8 --net_type=cnn --dev=gpu

Testing

To evaluate the model on the test dataset

python main.py test [--args=value]

This uses the same arguments as above, except --dev is not included as cpu is used for testing. It generates a log file at results/log-1.txt, which includes the original dialogues with the true labels and the model predictions.

Currently there is a bug that shuffles the model predictions across dialogues. So to go around it, make sure to use a batch size of 1, i.e. --batch_size=1.