KAT-TSLF

Source code of paper "A Novel Three-Stage Learning Framework for Low-Resource Knowledge-Grounded Dialogue Generation".

Environments

• python 3.7
• transformers 4.2.2
• NLTK
• pytorch
• language_evaluation (install from SKT project)

Datasets and Models

1. Download Wizard-of-Wikipedia, CMU_DoG and pseudo dataset (used in stage II) from [OneDrive] [189 Clond] [Google Drive].
2. Download BART pre-trained on Reddit Conversation Corpus and Wikipedia dumps from [Google Drive] and [Google Drive].
3. (Optional) Download pre-trained checkpoint from [OneDrive] [189 Clond] [Google Drive].

(Optional) Run Stage II

bash scripts/warmup.sh

Run Stage III

Wizard-of-Wikipedia:

bash scripts/wizard.sh

CMU_DoG:

bash scripts/cmudog.sh

Low-resource on Wizard-of-Wikipedia:

bash scripts/wizardlr.sh

Zero-resource on Wizard-of-Wikipedia:

bash scripts/zr_wizard.sh

Low-resource on CMU_DoG:

bash scripts/cmudoglr.sh

Zero-resource on CMU_DoG:

bash scripts/zr_cmudog.sh

(Please adjust beam size appropriately)

Cite

@inproceedings{liu-etal-2021-three,
    title = "{A} {T}hree-{S}tage {L}earning {F}ramework for {L}ow-{R}esource {K}nowledge-{G}rounded {D}ialogue {G}eneration",
    author = "Liu, Shilei  and
      Zhao, Xiaofeng  and
      Li, Bochao  and
      Ren, Feiliang  and
      Zhang, Longhui  and
      Yin, Shujuan",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.173",
    pages = "2262--2272",
}