DialoGLUE
DialoGLUE is a conversational AI benchmark designed to encourage dialogue research in representation-based transfer, domain adaptation, and sample-efficient task learning. For a more detailed write-up of the benchmark check out our paper.
This repository contains all code related to the benchmark, including scripts for downloading relevant datasets, preprocessing them in a consistent format for benchmark submissions, evaluating any submission outputs, and running baseline models from the original benchmark description.
This repository also contains code for our NAACL paper, Example-Driven Intent Prediction with Observers.
Datasets
This benchmark is created by scripts that pull data from previously-published data resources. Thank you to the authors of those works for their great contributions:
| Dataset | Size | Description | License |
|---|---|---|---|
| Banking77 | 13K | online banking queries | CC-BY-4.0 |
| HWU64 | 11K | popular personal assistant queries | CC-BY-SA 3.0 |
| CLINC150 | 20K | popular personal assistant queries | CC-BY-SA 3.0 |
| Restaurant8k | 8.2K | restaurant booking domain queries | CC-BY-4.0 |
| DSTC8 SGD | 20K | multi-domain, task-oriented conversations between a human and a virtual assistant | CC-BY-SA 4.0 International |
| TOP | 45K | compositional queries for hierachical semantic representations | CC-BY-SA |
| MultiWOZ 2.1 | 12K | multi-domain dialogues with multiple turns | MIT |
Data Download
To download/process the various datasets that are part of the DialoGLUE benchmark, run bash download_data.sh from data_utils.
Upon completion, your DialoGLUE folder should contain the following:
dialoglue/
βββ banking
β βββ categories.json
β βββ test.csv
β βββ train_10.csv
β βββ train_5.csv
β βββ train.csv
β βββ val.csv
βββ clinc
β βββ categories.json
β βββ test.csv
β βββ train_10.csv
β βββ train_5.csv
β βββ train.csv
β βββ val.csv
βββ dstc8_sgd
β βββ stats.csv
β βββ test.json
β βββ train_10.json
β βββ train.json
β βββ val.json
β βββ vocab.txt
βββ hwu
β βββ categories.json
β βββ test.csv
β βββ train_10.csv
β βββ train_5.csv
β βββ train.csv
β βββ val.csv
βββ mlm_all.txt
βββ multiwoz
β βββ MULTIWOZ2.1
β β βββ dialogue_acts.json
β β βββ README.txt
β β βββ test_dials.json
β β βββ train_dials.json
β β βββ val_dials.json
β βββ MULTIWOZ2.1_fewshot
β βββ dialogue_acts.json
β βββ README.txt
β βββ test_dials.json
β βββ train_dials.json
β βββ val_dials.json
βββ restaurant8k
β βββ test.json
β βββ train_10.json
β βββ train.json
β βββ val.json
β βββ vocab.txt
βββ top
βββ eval.txt
βββ test.txt
βββ train_10.txt
βββ train.txt
βββ vocab.intent
βββ vocab.slot
The files with a _10 suffix (e.g., banking/train_10.csv) are used for the few-shot experiments, wherein models are trained with only 10% of the datasets.
EvalAI Leaderboard
The DialoGLUE benchmark is hosted on EvalAI and we invite submissions to our leaderboard. The submission should be a JSON file with keys corresponding to each of the seven DialoGLUE datasets:
{"banking": [*banking outputs*], "hwu": [*hwu outputs*], ..., "multiwoz": [*multiwoz outputs*]}
Given a set of seven model checkpoints, you can edit and run dump_outputs.py to generate a valid submission file. For the intent classification tasks (HWU64, Banking77, CLINC150), the outputs are a list of intent classes. For the slot filling tasks (Restaurant8k, DSTC8), the outputs are a list of spans. For the TOP dataset, the outputs are a list of (intent, slots) pairs wherein each slot is the path from the root to the leaf node. For MultiWOZ, the output follows the TripPy format and corresponds to the pred_res.test.final.json output file. We strongly recommend using the dump_outputs.py script to generate outputs.
For the few-shot experimental setting, we only train models on a subset (roughly 10%) of the training data. The specific data splits are produced by running download_data.sh. To mitigate the impact of random initialization, we ask that you train 5 models for each of the few-shot tasks and submit the output of all 5 models. The scores on the leaderboard will be the average of these five runs.
The few-shot submission file format is as follows:
{"banking": [*banking outputs from model 1*, ... *banking outputs from model 5*], ...}
You may run dump_outputs_fewshot.py to generate a valid submission file given the model paths corresponding to all of the runs.
Training
Almost all of the models can be trained/evaluated using the run.py script. MultiWOZ is the exception, and relies on the modified open-sourced TripPy implementation.
The commands for training/evaluating models are as follows. If you want to only run inference/evaluation, simply change --num_epochs to 0.
To train using example-driven intent prediction, add the --example flag to the training script. To use observer nodes, add the --use_observers flag.
Checkpoints
The relevant convbert and convbert-dg models can be found here.
HWU64
python run.py \
--train_data_path data_utils/dialoglue/hwu/train.csv \
--val_data_path data_utils/dialoglue/hwu/val.csv \
--test_data_path data_utils/dialoglue/hwu/test.csv \
--token_vocab_path bert-base-uncased-vocab.txt \
--train_batch_size 64 --dropout 0.1 --num_epochs 100 --learning_rate 6e-5 \
--model_name_or_path convbert-dg --task intent --do_lowercase --max_seq_length 50 --mlm_pre --mlm_during --dump_outputs \
Banking77
python run.py \
--train_data_path data_utils/dialoglue/banking/train.csv \
--val_data_path data_utils/dialoglue/banking/val.csv \
--test_data_path data_utils/dialoglue/banking/test.csv \
--token_vocab_path bert-base-uncased-vocab.txt \
--train_batch_size 32 --grad_accum 2 --dropout 0.1 --num_epochs 100 --learning_rate 6e-5 \
--model_name_or_path convbert-dg --task intent --do_lowercase --max_seq_length 100 --mlm_pre --mlm_during --dump_outputs \
CLINC150
python run.py \
--train_data_path data_utils/dialoglue/clinc/train.csv \
--val_data_path data_utils/dialoglue/clinc/val.csv \
--test_data_path data_utils/dialoglue/clinc/test.csv \
--token_vocab_path bert-base-uncased-vocab.txt \
--train_batch_size 64 --dropout 0.1 --num_epochs 100 --learning_rate 6e-5 \
--model_name_or_path convbert-dg --task intent --do_lowercase --max_seq_length 50 --mlm_pre --mlm_during --dump_outputs \
Restaurant8k
python run.py \
--train_data_path data_utils/dialoglue/restaurant8k/train.json \
--val_data_path data_utils/dialoglue/restaurant8k/val.json \
--test_data_path data_utils/dialoglue/restaurant8k/test.json \
--token_vocab_path bert-base-uncased-vocab.txt \
--train_batch_size 64 --dropout 0.1 --num_epochs 100 --learning_rate 6e-5 \
--model_name_or_path convbert-dg --task slot --do_lowercase --max_seq_length 50 --mlm_pre --mlm_during --dump_outputs \
DSTC8
python run.py \
--train_data_path data_utils/dialoglue/dstc8/train.json \
--val_data_path data_utils/dialoglue/dstc8/val.json \
--test_data_path data_utils/dialoglue/dstc8/test.json \
--token_vocab_path bert-base-uncased-vocab.txt \
--train_batch_size 64 --dropout 0.1 --num_epochs 100 --learning_rate 6e-5 \
--model_name_or_path convbert-dg --task slot --do_lowercase --max_seq_length 50 --mlm_pre --mlm_during --dump_outputs \
TOP
python run.py \
--train_data_path data_utils/dialoglue/top/train.txt \
--val_data_path data_utils/dialoglue/top/eval.txt \
--test_data_path data_utils/dialoglue/top/test.txt \
--token_vocab_path bert-base-uncased-vocab.txt \
--train_batch_size 64 --dropout 0.1 --num_epochs 100 --learning_rate 6e-5 \
--model_name_or_path convbert-dg --task top --do_lowercase --max_seq_length 50 --mlm_pre --mlm_during --dump_outputs \
MultiWOZ
The MultiWOZ code builds on the open-sourced TripPy implementation. To train/evaluate the model using our modifications (i.e., MLM pre-training), you can use trippy/DO.example.advanced.
Checkpoints
Checkpoints are released for (1) ConvBERT, (2) BERT-DG and (3) ConvBERT-DG. Given these pre-trained models and the code in this repo, all of our results can be reproduced.
Requirements
This project has been tested and is functional with Python 3.7. The Python dependencies are listed in the requirements.txt.
License
This project is licensed under the Apache-2.0 License.
Citation
If using these scripts or the DialoGLUE benchmark, please cite the following in any relevant work:
@article{MehriDialoGLUE2020,
title={DialoGLUE: A Natural Language Understanding Benchmark for Task-Oriented Dialogue},
author={S. Mehri and M. Eric and D. Hakkani-Tur},
journal={ArXiv},
year={2020},
volume={abs/2009.13570}
}If you use any code pertaining to example-driven training or observers, please cite the following paper:
@article{mehri2020example,
title={Example-Driven Intent Prediction with Observers},
author={Mehri, Shikib and Eric, Mihail and Hakkani-Tur, Dilek},
journal={arXiv preprint arXiv:2010.08684},
year={2020}
}