ConvLab is an open-source multi-domain end-to-end dialog system platform, aiming to enable researchers to quickly set up experiments with reusable components and compare a large set of different approaches, ranging from conventional pipeline systems to end-to-end neural models, in common environments.

ConvLab requires Python 3.6.5 or later. Windows is currently not offically supported.

Connda can be used set up a virtual environment with the version of Python required for ConvLab. If you already have a Python 3.6 or 3.7 environment you want to use, you can skip to the 'installing via pip' section.

1. Download and install Conda.

2. Create a Conda environment with Python 3.6.5

   conda create -n convlab python=3.6.5

3. Activate the Conda environment. You will need to activate the Conda environment in each terminal in which you want to use ConvLab.

   source activate convlab

Installing the library and dependencies is simple using pip.

pip install -r requirements.txt

Docker provides more isolation and consistency, and also makes it easy to distribute your environment to a compute cluster.

Once you have installed Docker just run the following commands to get an environment that will run on either the cpu or gpu.

1. Pull docker
   docker pull convlab/convlab:0.1

2. Run docker
   docker run -it --rm convlab/convlab:0.1

Once you've downloaded ConvLab and installed required packages, you can run the command-line interface with the python run.py command.

$ python run.py {spec file} {spec name} {mode}

For example:

# to evaluate a dialog system consisting of NLU(OneNet), DST(Rule), Policy(Rule), NLG(Template) on the MultiWOZ environment
$ python run.py demo.json onenet_rule_rule_template eval

# to see natural language utterances 
$ LOG_LEVEL=NL python run.py demo.json onenet_rule_rule_template eval

# to see natural language utterances and dialog acts 
$ LOG_LEVEL=ACT python run.py demo.json onenet_rule_rule_template eval

# to see natural language utterances, dialog acts and state representation
$ LOG_LEVEL=STATE python run.py demo.json onenet_rule_rule_template eval

# to train a DQN policy with NLU(OneNet), DST(Rule), NLG(Template) on the MultiWOZ environment
$ python run.py demo.json onenet_rule_dqn_template train

# to use the policy trained above
$ python run.py output/onenet_rule_dqn_template_{timestamp}/onenet_rule_dqn_template_spec.json onenet_rule_dqn_template eval@onenet_rule_dqn_template_t0_s0

Note that currently ConvLab can only train the policy component by interacting with a user simulator. For other components, ConvLab supports offline supervise learning. For example, you can train a NLU model using the local training script as in OneNet.

A spec file is used to fully specify experiments including a dialog agent and a user simulator. It is a JSON of multiple experiment specs, each containing the keys agent, env, body, meta, search.

We based our implementation on SLM-Lab. For an introduction to these concepts, you should check these docs.

Instead of writing one from scratch, you are welcome to modify the convlab/spec/demo.json file. Once you have created a new spec file, place it under convlab/spec directory and run your experiments. Note that you don't have to prepend convlab/spec/ before your spec file name.

The ConvLab team welcomes contributions from the community. Pull requests must have one approving review and no requested changes before they are merged. The ConvLab team reserve the right to reject or revert contributions that we don't think are good additions.

If you use ConvLab in your research, please cite ConvLab: Multi-Domain End-to-End Dialog System Platform.

@inproceedings{lee2019convlab,
  title={ConvLab: Multi-Domain End-to-End Dialog System Platform},
  author={Lee, Sungjin and Zhu, Qi and Takanobu, Ryuichi and Li, Xiang and Zhang, Yaoqin and Zhang, Zheng and Li, Jinchao and Peng, Baolin and Li, Xiujun and Huang, Minlie and others},
  booktitle={Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},
  year={2019}
}