DukeNet (SIGIR 2020 full paper)

The code for DukeNet: A Dual Knowledge Interaction Network for Knowledge-Grounded Conversation. This work was partly supported by the Tencent AI Lab Rhino-Bird Focused Research Program (JR201932).

Reference

If you use any source code included in this repo in your work, please cite the following paper.

@inproceedings{chuanmeng2020dukenet,
author = {Meng, Chuan and Ren, Pengjie and Chen, Zhumin and Sun, Weiwei and Ren, Zhaochun and Tu, Zhaopeng and Rijke, Maarten de},
title = {DukeNet: A Dual Knowledge Interaction Network for Knowledge-Grounded Conversation},
year = {2020},
booktitle = {SIGIR},
pages = {1151–1160},
}

Requirements

  • python 3.6
  • pytorch 1.2.0-1.4.0
  • transformers 2.6

Datasets

We use Wizard of Wikipedia and Holl-E datasets. Note that we used modified verion of Holl-E relased by Kim et al (But they don't release the validation set).
Both datasets have already been processed into our defined format, which could be directly used by our model.

You can manually download the datasets at here, and please create folder datasets in the root directory and put the files in it.

Running Codes

Note that it's rather time-consuming to train the DukeNet with BERT and dual learning. Therefore we upload our pretrained checkpoints on two datasets, and you can manually download them at here.

Please create folder output in the root directory and put the files in it.

Using pretrained checkpoints

To directly execute inference process on Wizard of Wikipedia dataset, run:

python DukeNet/Run.py --name DukeNet_WoW --dataset wizard_of_wikipedia --mode inference

Wizard of Wikipedia (test seen)
{"F1": 19.33,
 "BLEU-1": 17.99,
 "BLEU-2": 7.51,
 "BLEU-3": 4.04,
 "BLEU-4": 2.46,
 "ROUGE_1_F1": 25.38,
 "ROUGE_2_F1": 6.83,
 "ROUGE_L_F1": 18.67,
 "METEOR": 17.23,
 "t_acc": 81.9,
 "s_acc": 25.58}

Wizard of wikipedia (test unseen)
{"F1": 17.09,
 "BLEU-1": 16.34,
 "BLEU-2": 5.99,
 "BLEU-3": 2.85,
 "BLEU-4": 1.69,
 "ROUGE_1_F1": 23.45,
 "ROUGE_2_F1": 5.31,
 "ROUGE_L_F1": 16.95,
 "METEOR": 15.22,
 "t_acc": 75.88,
 "s_acc": 20.07}

Note: The inference process will create test_seen_result.json(test_unseen_result.json) and test_seen_xx.txt(test_unseen_xx.txt) in the directory output/DukeNet_WoW, where the former records the model performance on automatic evaluation metrics for different epochs, and the latter records the response generated by the model.

To directly execute inference process on Holl-E dataset, run:

python DukeNet/Run.py --name DukeNet_Holl_E --dataset holl_e --mode inference

Holl-E (single golden reference)
 {"F1": 30.58,
  "BLEU-1": 30.11,
  "BLEU-2": 22.51,
  "BLEU-3": 20.42,
  "BLEU-4": 19.43,
  "ROUGE_1_F1": 36.67,
  "ROUGE_2_F1": 23.24,
  "ROUGE_L_F1": 31.68,
  "METEOR": 31.27,
  "t_acc": 92.79,
  "s_acc": 30.4}

If you want to get the results in the setting of multiple golden references, run:

python DukeNet/CumulativeTrainer.py 

Holl-E (multiple golden references)
{"F1": 37.68,
 "BLEU-1": 40.43, 
 "BLEU-2": 31.03, 
 "BLEU-3": 28.21, 
 "BLEU-4": 26.97, 
 "ROUGE_1_F1": 43.36, 
 "ROUGE_2_F1": 30.42, 
 "ROUGE_L_F1": 38.28, 
 "METEOR": 38.14,  
 "s_acc": 40.81}

Retraining

To execute warm-up training phase and dual interaction training phase, run sequentially:

Wizard of Wikipedia
python DukeNet/Run.py --name DukeNet_WoW --dataset wizard_of_wikipedia --mode train
python DukeNet/Dual_Run.py --name DukeNet_WoW --dataset wizard_of_wikipedia 

Holl-E
python DukeNet/Run.py --name DukeNet_Holl_E --dataset holl_e --mode train
python DukeNet/Dual_Run.py --name DukeNet_Holl_E --dataset holl_e

GitHub