Skip to content

Latest commit

 

History

History
 
 

Dialogue-PLATO

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

PLATO

PLATO: Pre-trained Dialogue Generation Model with Discrete Latent Variable paper link

***** Update *****

Jul. 9, 2020: We are opening PLATO-2, a large-scale generative model with latent space for open-domain dialogue systems.

Nov. 14, 2019: Support new APIs in paddlepaddle 1.6 (model files in the link have been updated accordingly), multi-GPU training and decoding strategy of top-k sampling. Release our baseline model PLATO w/o latent.

Requirements

- python >= 3.6
- paddlepaddle == 1.6.1
- numpy
- nltk
- tqdm
- visualdl >= 1.3.0 (optional)
- regex

Recommend you install to python packages by command: pip install -r requirement.txt

Pre-trained dialogue generation model

A novel pre-training model for dialogue generation is introduced in this work, incorporated with latent discrete variables for one-to-many relationship modeling. Our model is flexible enough to support various kinds of conversations, including chit-chat, knowledge grounded dialogues, and conversational question answering. The pre-training is carried out with Reddit and Twitter corpora. You can download the uncased pre-trained model from:

  • PLATO, uncased model: 12-layers, 768-hidden, 12-heads, 132M parameters
  • PLATO w/o latent, uncased model: 12-layers 768-hidden, 12-heads, 109M parameters
mv /path/to/model.tar.gz .
tar xzf model.tar.gz

Fine-tuning

We also provide instructions to fine-tune PLATO on different conversation datasets (chit-chat, knowledge grounded dialogues and conversational question answering).

Data preparation

Download data from the link. The tar file contains three processed datasets: DailyDialog, PersonaChat and DSTC7_AVSD.

mv /path/to/data.tar.gz .
tar xzf data.tar.gz

Data format

Our model supports two kinds of data formats for dialogue context: multi and multi_knowledge.

  • multi: multi-turn dialogue context.
u_1 __eou__ u_2 __eou__ ... u_n \t r
  • multi_knowledge: multi-turn dialogue context with background knowledges.
k_1 __eou__ k_2 __eou__ ... k_m \t u_1 __eou__ u_2 __eou__ ... u_n \t r

If you want to use this model on other datasets, you can process your data accordingly.

Train

Fine-tuning the pre-trained model on different ${DATASET}.

# DailyDialog / PersonaChat / DSTC7_AVSD
DATASET=DailyDialog
sh scripts/${DATASET}/train.sh

After training, you can find the output folder outputs/${DATASET} (by default). It contatins best.model (best results on validation dataset), hparams.json (hyper-parameters of training script) and trainer.log (training log).

Fine-tuning the pre-trained model on multiple GPUs.

Note: You need to install NCCL library and set up the environment variable LD_LIBRARY properly.

sh scripts/DailyDialog/multi_gpu_train.sh

You can fine-tune PLATO w/o latent on different ${DATASET}. We provide an example script on DailyDialog dataset.

sh scripts/DailyDialog/baseline_train.sh

Recommended settings

For the fine-tuning of our pre-trained model, it usually requires about 10 epochs to reach convergence with learning rate = 1e-5 and about 2-3 epochs to reach convergence with learning rate = 5e-5.

GPU Memory batch size max len
16G 6 256
32G 12 256

Infer

Running inference on test dataset.

# DailyDialog / PersonaChat / DSTC7_AVSD
DATASET=DailyDialog
sh scripts/${DATASET}/infer.sh

# Running inference of PLATO w/o latent
sh scripts/DailyDialog/baseline_infer.sh

After inference, you can find the output foler outputs/${DATASET}.infer (by default). It contains infer_0.result.json (the inference result), hparams.json (hyper-parameters of inference scipt) and trainer.log (inference log).

If you want to use top-k sampling (beam search by default), you can follow the example script:

sh scripts/DailyDialog/topk_infer.sh

Result

DailyDialog

Model BLEU-1/2 Distinct-1/2 Fluency Coherence Informativeness Overall
Seq2Seq 0.336/0.268 0.030/0.128 1.85 0.37 0.44 0.33
iVAE_MI 0.309/0.249 0.029/0.250 1.53 0.34 0.59 0.30
Our w/o Latent 0.405/0.322 0.046/0.246 1.91 1.58 1.03 1.44
Our Method 0.397/0.311 0.053/0.291 1.97 1.57 1.23 1.48

PersonaChat

Model BLEU-1/2 Distinct-1/2 Knowledge R/P/F1 Fluency Coherence Informativeness Overall
Seq2Seq 0.448/0.353 0.004/0.016 0.004/0.016/0.006 1.82 0.37 0.85 0.34
LIC 0.405/0.320 0.019/0.113 0.042/0.154/0.064 1.95 1.34 1.09 1.29
Our w/o Latent 0.458/0.357 0.012/0.064 0.085/0.263/0.125 1.98 1.36 1.04 1.30
Our Method 0.406/0.315 0.021/0.121 0.142/0.461/0.211 1.99 1.51 1.70 1.50

DSTC7_AVSD

Model BELU-1 BELU-2 BLEU-3 BLEU-4 METEOR ROUGH-L CIDEr
Baseline 0.629 0.485 0.383 0.309 0.215 0.487 0.746
CMU 0.718 0.584 0.478 0.394 0.267 0.563 1.094
Our Method 0.784 0.637 0.525 0.435 0.286 0.596 1.209
Our Method Upper Bound 0.925 0.843 0.767 0.689 0.361 0.731 1.716

Note: In the experiments on DSTC7_AVSD, the response selection of our method is strengthened with an extra ranking step, which ranks the candidates according to the automatic scores and selects the top one as the final answer.

Citation

If you find PLATO useful in your work, please cite the following paper:

@inproceedings{bao2019plato,
    title={PLATO: Pre-trained Dialogue Generation Model with Discrete Latent Variable},
    author={Bao, Siqi and He, Huang and Wang, Fan and Wu, Hua and Wang, Haifeng},
    booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics},
    pages={85--96},
    year={2020}
}

Disclaimer

This project aims to facilitate further research progress in dialogue generation. Baidu is not responsible for the 3rd party's generation with the pre-trained system.

Contact information

For help or issues using PLATO, please submit a GitHub issue.

For personal communication related to PLATO, please contact Siqi Bao ([email protected]), or Huang He ([email protected]).