COLING2022-TAKE: Topic-shift Aware Knowledge sElection for Dialogue Generation

This repository contains the code for COLING-2022 full paper:TAKE: Topic-shift Aware Knowledge sElection for Dialogue Generation.

Please contact Chenxu Yang ([email protected]) if you have any question.

Requirements

• transformers==4.15.0
• python 3.9
• pytorch 1.10.1

Datasets

We use the Wizard of Wikipedia datasets preprocessed by Meng et al. You can download the datasets from here. After downloading, please put the files in the following 3 dirs:

./knowSelect/datasets/wizard_of_wikipedia/

./dialogen/datasets/wizard_of_wikipedia/

./dialogen/datasets/wow_gpt2/

Running TAKE Codes

Inference

We provide pretrained checkpoints to save your time, and you can acquire them here. You need to download the corresponding checkpoints and put them in the folder ./knowSelect/output/TAKE_WoW/model/ (KS) or the folder ./dialogen/output/TAKE_WoW/model/ (DG) according to their name. After that, you can run the infer_bash.sh in in the root directory to get the evaluation results.

Retraining

Of course, you can also retrain and evaluate TAKE by running the train_bash.sh in the root directory.

Citation

Please cite our paper if you use use source code of TAKE in your work:

Chenxu Yang, Zheng Lin, Jiangnan Li, Fandong Meng, Weiping Wang, Lanrui Wang, and Jie Zhou. 2022. TAKE: Topic-shift Aware Knowledge sElection for Dialogue Generation. In Proceedings of the 29th International Conference on Computational Linguistics, pages 253–265, Gyeongju, Republic of Korea. International Committee on Computational Linguistics.