Home

Awesome

ConvLab-3

PyPI PyPI - Python Version GitHub

ConvLab-3 is a flexible dialog system platform based on a unified data format for task-oriented dialog (TOD) datasets. The unified format serves as the adapter between TOD datasets and models: datasets are first transformed to the unified format and then loaded by models. In this way, the cost of adapting $M$ models to $N$ datasets is reduced from $M\times N$ to $M+N$. While retaining all features of ConvLab-2, ConvLab-3 greatly enlarges supported datasets and models thanks to the unified format, and enhances the utility of reinforcement learning (RL) toolkit for dialog policy module. For typical usage, see our paper. Datasets and Trained models are also available on Hugging Face Hub.

Updates

Installation

You can install ConvLab-3 in one of the following ways according to your need. We use torch>=1.10.1,<=1.13 and transformers>=4.17.0,<=4.24.0. Higher versions of torch and transformers may also work.

Git clone and pip install in development mode (Recommend)

For the latest and most configurable version, we recommend installing ConvLab-3 in development mode.

Clone the newest repository:

git clone --depth 1 https://github.com/ConvLab/ConvLab-3.git

Install ConvLab-3 via pip:

cd ConvLab-3
pip install -e .

Pip install from PyPI

To use ConvLab-3 as an off-the-shelf tool, you can install via:

pip install convlab

Note that the data directory will not be included due to the package size limitation.

Using Docker

We also provide Dockerfile for building docker. Basically it uses the requirement.txt and then installs ConvLab-3 in development mode.

# create image
docker build -t convlab .

# run container
docker run -dit convlab

# open bash in container
docker exec -it CONTAINER_ID bash

Tutorials

Unified Datasets

Current datasets in unified data format: (DA-U/DA-S stands for user/system dialog acts)

DatasetDialogsGoalDA-UDA-SStateAPI resultDataBase
Camrest676:white_check_mark::white_check_mark::white_check_mark::white_check_mark::white_check_mark:
WOZ 2.01200:white_check_mark::white_check_mark:
KVRET3030:white_check_mark::white_check_mark::white_check_mark:
DailyDialog13118:white_check_mark:
Taskmaster-113175:white_check_mark::white_check_mark::white_check_mark:
Taskmaster-217303:white_check_mark::white_check_mark::white_check_mark:
MultiWOZ 2.110438:white_check_mark::white_check_mark::white_check_mark::white_check_mark::white_check_mark:
Schema-Guided22825:white_check_mark::white_check_mark::white_check_mark::white_check_mark:
MetaLWOZ40203:white_check_mark:
CrossWOZ (zh)6012:white_check_mark::white_check_mark::white_check_mark::white_check_mark::white_check_mark::white_check_mark:
Taskmaster-323757:white_check_mark::white_check_mark::white_check_mark::white_check_mark:

Unified datasets are available under data/unified_datasets directory as well as Hugging Face Hub. We will continue adding more datasets listed in this issue. If you want to add a listed/custom dataset to ConvLab-3, you can create an issue for discussion and then create pull-request. We will list you as the contributors and highly appreciate your contributions!

Models

We list newly integrated models in ConvLab-3 that support unified data format and obtain strong performance. You can follow the link for more details about these models. Other models can be used in the same way as in ConvLab-2.

TaskModelsInputOutput
Response GenerationT5RG, LLMsContextResponse
Goal-to-DialogueT5Goal2DialogueGoalDialog
Natural Language UnderstandingT5NLU, BERTNLU, MILU, LLMsContextDA-U
Dialog State TrackingT5DST, SUMBT, SetSUMBT, TripPy, LLMsContextState
RL PolicyDDPT, PPO, PGState, DA-U, DBDA-S
Word-PolicyLAVAContext, State, DBResponse
Natural Language GenerationT5NLG, SC-GPT, LLMsDA-SResponse
End-to-EndSOLOISTContext, DBState, Response
User simulatorTUS, GenTUS, LLMsGoal, DA-SDA-U, (Response)

Trained models are available on Hugging Face Hub.

Contributing

We welcome contributions from community. Please see issues to find what we need.

Code Structure

.
├── convlab                             # Source code, installed in pypi package
│   ├── dialog_agent                    # Interface for dialog agent and session
│   ├── base_models
│   │   ├── llm                         # LLM-based models
│   │   │   ├── user_simulator          # LLM-based user simulator and RG
│   │   │   ├── dst                     # LLM-based DST
│   │   │   ├── nlu                     # LLM-based NLU
│   │   │   └── nlg                     # LLM-based NLG
│   │   │
│   │   └── t5                          # T5 models with a unified training script
│   │       ├── goal2dialogue           # T5Goal2Dialogue
│   │       ├── dst                     # T5DST
│   │       ├── nlu                     # T5NLU
│   │       ├── nlg                     # T5NLG
│   │       └── rg                      # T5RG
│   │
│   ├── nlu                             # NLU models, interface, and evaluation script
│   │   ├── jointBERT                   # BERTNLU
│   │   ├── milu                        # MILU
│   │   └── svm                         # SVMNLU*
│   │
│   ├── laug                            # Language understanding AUGmentation (LAUG) toolkit
│   │
│   ├── dst                             # DST models, interface, and evaluation script
│   │   ├── rule                        # RuleDST
│   │   ├── setsumbt                    # SetSUMBT, has uncertainty estimates
│   │   ├── sumbt                       # SUMBT
│   │   ├── trippy                      # TripPy
│   │   ├── trade                       # TRADE*
│   │   ├── comer                       # COMER*
│   │   ├── mdbt                        # MDBT*
│   │   └── dstc9                       # scripts for DSTC9 cross-lingual DST evaluation
│   │
│   ├── policy                          # Policy models, interface, and RL toolkit
│   │   ├── vector                      # vectorizer class
│   │   ├── plot_results                # RL plotting tool
│   │   ├── mle                         # MLE (imitation learning) policy
│   │   ├── pg                          # Policy Gradient
│   │   ├── ppo                         # Proximal Policy Optimization
│   │   ├── vtrace_DPT                  # DDPT
│   │   ├── lava                        # LAVA
│   │   ├── rule                        # Rule policies and rule-based user simulators 
│   │   ├── tus                         # TUS
│   │   ├── genTUS                      # GenTUS
│   │   ├── dqn                         # DQN*
│   │   ├── gdpl                        # GDPL*
│   │   ├── vhus                        # VHUS*
│   │   ├── hdsa                        # HDSA*
│   │   ├── larl                        # LARL*
│   │   └── mdrg                        # MDRG*
│   │
│   ├── nlg                             # NLG models, interface, and evaluation script
│   │   ├── scgpt                       # SC-GPT
│   │   ├── sclstm                      # SC-LSTM
│   │   └── template                    # TemplateNLG*
│   │
│   ├── e2e                             # End2End models
│   │   ├── soloist                     # SOLOIST
│   │   ├── damd                        # DAMD*
│   │   └── sequicity                   # Sequicity*
│   │
│   ├── evaluator                       # Evaluator for interactive evaluation
│   ├── human_eval                      # Human evaluation with AMT
│   ├── task                            # Goal generators for MultiWOZ, CrossWOZ, and Camrest
│   ├── util
│   │   └── unified_datasets_util.py    # Utility function for unified data format
│   └── deploy                          # Deploy system for human conversion
│
├── data                                # Data dir, not included in pypi package
│   ├── ...                             # ConvLab-2 data, not available for pypi installation
│   └── unified_datasets                # Unified datasets, available for pypi installation
├── examples
│   └── agent_examples                  # Examples of building user and system agents
└── tutorials                           # Tutorials

*: models do not support unified datasets, only support MultiWOZ.

Team

ConvLab-3 is maintained and developed by Tsinghua University Conversational AI group (THU-COAI), the Dialogue Systems and Machine Learning Group at Heinrich Heine University, Düsseldorf, Germany and Microsoft Research (MSR).

We would like to thank all contributors of ConvLab:

Yan Fang, Zhuoer Feng, Jianfeng Gao, Qihan Guo, Kaili Huang, Minlie Huang, Sungjin Lee, Bing Li, Jinchao Li, Xiang Li, Xiujun Li, Jiexi Liu, Lingxiao Luo, Wenchang Ma, Mehrad Moradshahi, Baolin Peng, Runze Liang, Ryuichi Takanobu, Dazhen Wan, Hongru Wang, Jiaxin Wen, Yaoqin Zhang, Zheng Zhang, Qi Zhu, Xiaoyan Zhu, Carel van Niekerk, Christian Geishauser, Hsien-chin Lin, Nurul Lubis, Xiaochen Zhu, Michael Heck, Shutong Feng, Milica Gašić.

Citing

If you use ConvLab-3 in your research, please cite:

@article{zhu2022convlab3,
    title={ConvLab-3: A Flexible Dialogue System Toolkit Based on a Unified Data Format},
    author={Qi Zhu and Christian Geishauser and Hsien-chin Lin and Carel van Niekerk and Baolin Peng and Zheng Zhang and Michael Heck and Nurul Lubis and Dazhen Wan and Xiaochen Zhu and Jianfeng Gao and Milica Gašić and Minlie Huang},
    journal={arXiv preprint arXiv:2211.17148},
    year={2022},
    url={http://arxiv.org/abs/2211.17148}
}

License

Apache License 2.0