Official Implementation for Stylized Dialogue Generation with Feature-Guided Knowledge Augmentation (EMNLP2023 Findings).

Introducing KASDG: A Stylized Dialogue Response Generator that leverages style corpus from Knowledge Base perspective. KASDG extracts style knowledge from a style corpus and employs a uniquely designed Feature-Guided Selection Module, integrating Response-Related Contrastive Learning and Style Responsiveness Kullback-Leibler loss.

• Reddit : Dialogue Corpus.

• TCFC : Style corpus that contains the Formal and Informal styles.

• ArXiv & Holmes : Style corpus that contains Holmes and ArXiv styles.

We tested our code on CUDA 11.4.

cd KASDG
conda create -n kasdg python=3.10
pip install -r requirements.txt
conda activate kasdg

• Model: Download the pretrained weights of bart and bart_dialogue. Move them to model and name as bart,bart_dialogue, respectively.
• Dataset : We provide our used dataset in data. Same data can be downloaded from links mentioned in Dataset section.
• Preprocess : To preprocess the data and complete the retrieval process overhead python src/prepare_data.py --dataset='your_dataset' --preprocess=1

• To train a model python train.py --dataset='your_dataset'
• To predict or test with a trained model python eval.py --load_path='your_model_path' --load_step==your_step --dataset='your_dataset'

We use the same evaluation metrics as the original paper. For TCFC evaluations, please refer to this paper. For ArXiv&Holmes evaluations, please refer to this paper. You can also use your own metrics to evaluate the performance.

If you find our work useful in your research, please consider citing:

If you have any questions, please open a github issue or contact us:

Zekai Zhang: [email protected]

Jinpeng Li: [email protected]