RefNet (AAAI 2020 oral paper)

The code for RefNet: A Reference-aware Network for Background Based Conversation

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

@inproceedings{chuanmeng2020refnet,
 author = {Meng, Chuan and Ren, Pengjie and Chen, Zhumin and Monz, Christof and Ma, Jun and de Rijke, Maarten},
 booktitle = {Proceedings of the Thirty-Fourth AAAI Conference on Artificial Intelligence},
 title = {RefNet: A Reference-aware Network for Background Based Conversation},
 year = {2020}
}

Requirements

python 3.6
tensorflow-gpu 1.3

Run dataset preprocessing

First, you should download the raw data version of Holl-E, and put the raw data files (train_data.json, dev_data.json and test_data.json）in the directory /data. Then run the preprocessing script in the directory /data:

python preprocress.py

This will create the data in the setting of mixed-short background, which is the data version used in our paper. You can also change the setting in preprocress.py to create the data in the setting of oracle or mixed-long background.

Run training

To train your model, run:

python run.py --mode=train

This will create a directory log/RefNet/train/ where all checkpoints and other data will be saved. Then the model will start training using the train_*.bin files as training data.

Hyperparameter: If you want to change some hyperparameters, you can change some settings in the file config.yaml carefully.

Run validation

You may want to run a concurrent evaluation job, that runs your model on the validation set and logs the results of automatic evaluation (F1, BLEU-4, ROUGE-1, ROUGE-2 and ROUGE-L ). To do this, run:

python run.py --mode=val

This will create val_result.json and Validation_Infer_ckpt-xxxx in the directory log/RefNet/, where val_result.json records the model performance about automatic evaluation metrics on the validation set for every epoch, and Validation_Infer_ckpt-xxxx records the response outputted by our model.

Note: you can run validation after finishing the training, or at the same time with training.

Run testing

After finishing the proprocess of training and validation, you should open the file val_result.json to select the best model based on the validation results according to BLEU metric or other metrics you like. Next, you should copy the model file path, for examplelog/RefNet/train/model.ckpt-10775, and run testing on this model:

python run.py --mode=test --appoint_test log/RefNet/train/model.ckpt-10775

This will create test_result.json and Test_Infer_ckpt-10775 in the directory log/RefNet/, where the purpose of these two files is consistent with the validation process.

If you want to test the model in the setting of multiple golden references, run:

python run.py --mode=test --appoint_test log/RefNet/train/model.ckpt-10775 --multi_label_eval True