Skip to content

ConvLab-2: An Open-Source Toolkit for Building, Evaluating, and Diagnosing Dialogue Systems

License

Notifications You must be signed in to change notification settings

thu-coai/ConvLab-2

Repository files navigation

ConvLab-2

Build Status

ConvLab-2 is an open-source toolkit that enables researchers to build task-oriented dialogue systems with state-of-the-art models, perform an end-to-end evaluation, and diagnose the weakness of systems. As the successor of ConvLab, ConvLab-2 inherits ConvLab's framework but integrates more powerful dialogue models and supports more datasets. Besides, we have developed an analysis tool and an interactive tool to assist researchers in diagnosing dialogue systems. [paper]

Updates

2022.11.30:

  • ConvLab-3 [paper] release! Building dialog systems on custom datasets is easier now. Most part of ConvLab-2 is retained. New features include:
    • We proposed a unified format for TOD datasets, transformed many commonly used datasets, and adapted models to support the unified format, facilitating research involving many datasets.
    • We added powerful transformer-based models for every module, including two transferable user simulators which can be used for custom datasets.
    • We advanced the RL toolkit. We simplified the process of building the dialogue system and its RL environment, provided plotting tools to compare policies, and offered a wide range of evaluation metrics.

2022.11.14:

  • Due to the potential security risk, The trained models of ConvLab-2 hosted at Azure can not be accessed currently. Therefore, we copied these models and placed them in Hugging Face. We've replaced the model URL in the ConvLab-2 code with the model URL in our Hugging Face repo. If you try to use trained models of ConvLab-2, make sure to update the code.

2021.9.13:

  • Add MultiWOZ 2.3 dataset in data dir. The dataset adds co-reference annotations in addition to corrections of dialogue acts and dialogue states. [paper]

2021.6.18:

  • Add LAUG, an open-source toolkit for Language understanding AUGmentation. It is an automatic method to approximate the natural perturbations to existing data. Augmented data could be used to conduct black-box robustness testing or enhancing training. [paper]
  • Add SC-GPT for NLG. [paper]

Installation

Require python >= 3.6.

Clone this repository:

git clone https://github.com/thu-coai/ConvLab-2.git

Install ConvLab-2 via pip:

cd ConvLab-2
pip install -e .

Tutorials

Documents

Our documents are on https://thu-coai.github.io/ConvLab-2_docs/convlab2.html.

Models

We provide following models:

  • NLU: SVMNLU, MILU, BERTNLU
  • DST: rule, TRADE, SUMBT
  • Policy: rule, Imitation, REINFORCE, PPO, GDPL, MDRG, HDSA, LaRL
  • Simulator policy: Agenda, VHUS
  • NLG: Template, SCLSTM
  • End2End: Sequicity, DAMD, RNN_rollout

For more details about these models, You can refer to README.md under convlab2/$module/$model/$dataset dir such as convlab2/nlu/jointBERT/multiwoz/README.md.

Supported Datasets

  • Multiwoz 2.1
    • We add user dialogue act (inform, request, bye, greet, thank), remove 5 sessions that have incomplete dialogue act annotation and place it under data/multiwoz dir.
    • Train/val/test size: 8434/999/1000. Split as original data.
    • LICENSE: Attribution 4.0 International, url: http://creativecommons.org/licenses/by/4.0/
  • CrossWOZ
    • We offers a rule-based user simulator and a complete set of models for building a pipeline system on the CrossWOZ dataset. We correct few state annotation and place it under data/crosswoz dir.
    • Train/val/test size: 5012/500/500. Split as original data.
    • LICENSE: Attribution 4.0 International, url: http://creativecommons.org/licenses/by/4.0/
  • Camrest
    • We add system dialogue act (inform, request, nooffer) and place it under data/camrest dir.
    • Train/val/test size: 406/135/135. Split as original data.
    • LICENSE: Attribution 4.0 International, url: http://creativecommons.org/licenses/by/4.0/
  • Dealornot

End-to-end Performance on MultiWOZ

Notice: The results are for commits before bdc9dba (inclusive). We will update the results after improving user policy.

We perform end-to-end evaluation (1000 dialogues) on MultiWOZ using the user simulator below (a full example on tests/test_end2end.py) :

# BERT nlu trained on sys utterance
user_nlu = BERTNLU(mode='sys', config_file='multiwoz_sys_context.json', model_file='https://huggingface.co/ConvLab/ConvLab-2_models/resolve/main/bert_multiwoz_sys_context.zip')
user_dst = None
user_policy = RulePolicy(character='usr')
user_nlg = TemplateNLG(is_user=True)
user_agent = PipelineAgent(user_nlu, user_dst, user_policy, user_nlg, name='user')

analyzer = Analyzer(user_agent=user_agent, dataset='multiwoz')

set_seed(20200202)
analyzer.comprehensive_analyze(sys_agent=sys_agent, model_name='sys_agent', total_dialog=1000)

Main metrics (refer to convlab2/evaluator/multiwoz_eval.py for more details):

  • Complete: whether complete the goal. Judged by the Agenda policy instead of external evaluator.
  • Success: whether all user requests have been informed and the booked entities satisfy the constraints.
  • Book: how many the booked entities satisfy the user constraints.
  • Inform Precision/Recall/F1: how many user requests have been informed.
  • Turn(succ/all): average turn number for successful/all dialogues.

Performance (the first row is the default config for each module. Empty entries are set to default config.):

NLU DST Policy NLG Complete rate Success rate Book rate Inform P/R/F1 Turn(succ/all)
BERTNLU RuleDST RulePolicy TemplateNLG 90.5 81.3 91.1 79.7/92.6/83.5 11.6/12.3
MILU RuleDST RulePolicy TemplateNLG 93.3 81.8 93.0 80.4/94.7/84.8 11.3/12.1
BERTNLU RuleDST RulePolicy SCLSTM 48.5 40.2 56.9 62.3/62.5/58.7 11.9/27.1
BERTNLU RuleDST MLEPolicy TemplateNLG 42.7 35.9 17.6 62.8/69.8/62.9 12.1/24.1
BERTNLU RuleDST PGPolicy TemplateNLG 37.4 31.7 17.4 57.4/63.7/56.9 11.0/25.3
BERTNLU RuleDST PPOPolicy TemplateNLG 75.5 71.7 86.6 69.4/85.8/74.1 13.1/17.8
BERTNLU RuleDST GDPLPolicy TemplateNLG 49.4 38.4 20.1 64.5/73.8/65.6 11.5/21.3
None TRADE RulePolicy TemplateNLG 32.4 20.1 34.7 46.9/48.5/44.0 11.4/23.9
None SUMBT RulePolicy TemplateNLG 34.5 29.4 62.4 54.1/50.3/48.3 11.0/28.1
BERTNLU RuleDST MDRG None 21.6 17.8 31.2 39.9/36.3/34.8 15.6/30.5
BERTNLU RuleDST LaRL None 34.8 27.0 29.6 49.1/53.6/47.8 13.2/24.4
None SUMBT LaRL None 32.9 23.7 25.9 48.6/52.0/46.7 12.5/24.3
None None DAMD* None 39.5 34.3 51.4 60.4/59.8/56.3 15.8/29.8

*: end-to-end models used as sys_agent directly.

Module Performance on MultiWOZ

NLU

By running convlab2/nlu/evaluate.py MultiWOZ $model all:

Precision Recall F1
BERTNLU 82.48 85.59 84.01
MILU 80.29 83.63 81.92
SVMNLU 74.96 50.74 60.52

DST

By running convlab2/dst/evaluate.py MultiWOZ $model:

Joint accuracy Slot accuracy Joint F1
MDBT 0.06 0.89 0.43
SUMBT 0.30 0.96 0.83
TRADE 0.40 0.96 0.84

Policy

Notice: The results are for commits before bdc9dba (inclusive). We will update the results after improving user policy.

By running convlab2/policy/evalutate.py --model_name $model

Task Success Rate
MLE 0.56
PG 0.54
PPO 0.89
GDPL 0.58

NLG

By running convlab2/nlg/evaluate.py MultiWOZ $model sys

corpus BLEU-4
Template 0.3309
SCLSTM 0.4884

Translation-train SUMBT for cross-lingual DST

Train

With Convlab-2, you can train SUMBT on a machine-translated dataset like this:

# train.py
import os
from sys import argv

if __name__ == "__main__":
    if len(argv) != 2:
        print('usage: python3 train.py [dataset]')
        exit(1)
    assert argv[1] in ['multiwoz', 'crosswoz']

    from convlab2.dst.sumbt.multiwoz_zh.sumbt import SUMBT_PATH
    if argv[1] == 'multiwoz':
        from convlab2.dst.sumbt.multiwoz_zh.sumbt import SUMBTTracker as SUMBT
    elif argv[1] == 'crosswoz':
        from convlab2.dst.sumbt.crosswoz_en.sumbt import SUMBTTracker as SUMBT

    sumbt = SUMBT()
    sumbt.train(True)

Evaluate

Execute evaluate.py (under convlab2/dst/) with following command:

python3 evaluate.py [CrossWOZ-en|MultiWOZ-zh] [val|test|human_val]

evaluation of our pre-trained models are: (joint acc.)

type CrossWOZ-en MultiWOZ-zh
val 12.4% 48.5%
test 12.4% 46.0%
human_val 10.6% 47.4%

human_val option will make the model evaluate on the validation set translated by human.

Note: You may want to download pre-traiend BERT models and translation-train SUMBT models provided by us.

Without modifying any code, you could:

  • download pre-trained BERT models from:

    extract it to ./pre-trained-models.

  • for translation-train SUMBT model:

    • trained on CrossWOZ-en
    • trained on MultiWOZ-zh
    • Say the data set is CrossWOZ (English), (after extraction) just save the pre-trained model under ./convlab2/dst/sumbt/crosswoz_en/pre-trained and name it with pytorch_model.bin.

Issues

You are welcome to create an issue if you want to request a feature, report a bug or ask a general question.

Contributions

We welcome contributions from community.

  • If you want to make a big change, we recommend first creating an issue with your design.
  • Small contributions can be directly made by a pull request.
  • If you like make contributions to our library, see issues to find what we need.

Team

ConvLab-2 is maintained and developed by Tsinghua University Conversational AI group (THU-coai) and Microsoft Research (MSR).

We would like to thank:

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, Hongru Wang, Jiaxin Wen, Yaoqin Zhang, Zheng Zhang, Qi Zhu, Xiaoyan Zhu.

Citing

If you use ConvLab-2 in your research, please cite:

@inproceedings{zhu2020convlab2,
    title={ConvLab-2: An Open-Source Toolkit for Building, Evaluating, and Diagnosing Dialogue Systems},
    author={Qi Zhu and Zheng Zhang and Yan Fang and Xiang Li and Ryuichi Takanobu and Jinchao Li and Baolin Peng and Jianfeng Gao and Xiaoyan Zhu and Minlie Huang},
    year={2020},
    booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
}

@inproceedings{liu2021robustness,
    title={Robustness Testing of Language Understanding in Task-Oriented Dialog},
    author={Liu, Jiexi and Takanobu, Ryuichi and Wen, Jiaxin and Wan, Dazhen and Li, Hongguang and Nie, Weiran and Li, Cheng and Peng, Wei and Huang, Minlie},
    year={2021},
    booktitle={Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics},
}

License

Apache License 2.0