jq2276 / Learning2Copy

README

---

Codes for AAAI2021 paper ''Learning to Copy Coherent Knowledge for Response Generation.''

Requirements

• tqdm==4.54.0
• numpy==1.19.4
• nltk==3.5
• torch==1.7.1

Datasets

We use two datasets to implement our experiment, one is DuConv and the other is DuRecDial. Download and put the data under both data/resource/DuConv/ and data/resource/DuRecDial/, and rename them train/dev/test.txt under each folder. (e.g., /data/resource/DuConv/test.txt). To download the DuConv and DuRecDial datasets, please refer to the following papers:

DuConv:

Wu, W.; Guo, Z.; Zhou, X.; Wu, H.; Zhang, X.; Lian, R.; and Wang, H. 2019. Proactive Human-Machine Conversation with Explicit Conversation Goal. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 3794–3804.

DuRecDial:

Liu, Z.; Wang, H.; Niu, Z.-Y.; Wu, H.; Che, W.; and Liu, T. 2020. Towards Conversational Recommendation over Multi-Type Dialogs. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 1036–1049. Online: Association for Computational Linguistics.

Training & Testing

• We train our model on a single Nvidia Testla V100 machine. You can run bash run_train.sh to train the model with the default settings.
• After the training procedure, you can run bash run_test.sh to test the model.
• The training and testing data (DuConv or DuRecDial) can be changed through the argument data in both run_train.sh and run_test.sh.
• The hyperparameter settings can be found in network.py.