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]
2022.11.30:
2022.11.14:
2021.9.13:
data
dir. The dataset adds co-reference annotations in addition to corrections of dialogue acts and dialogue states. [paper]2021.6.18:
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 .
Our documents are on https://thu-coai.github.io/ConvLab-2_docs/convlab2.html.
We provide following models:
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
.
data/multiwoz
dir.data/crosswoz
dir.data/camrest
dir.data/dealornot
dir.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):
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.
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 |
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 |
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 |
By running convlab2/nlg/evaluate.py MultiWOZ $model sys
corpus BLEU-4 | |
---|---|
Template | 0.3309 |
SCLSTM | 0.4884 |
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)
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:
./convlab2/dst/sumbt/crosswoz_en/pre-trained
and name it with pytorch_model.bin
. You are welcome to create an issue if you want to request a feature, report a bug or ask a general question.
We welcome contributions from community.
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.
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},
}
Apache License 2.0