Sequential Knowledge Transformer (SKT)

This project hosts the code and dataset for our paper.

• Byeongchang Kim, Jaewoo Ahn and Gunhee Kim. Sequential Latent Knowledge Selection for Knowledge-Grounded Dialogue. In ICLR (spotlight), 2020. [OpenReview]

TL;DR: We propose a novel model named sequential knowledge transformer (SKT). To the best of our knowledge, our model is the first attempt to leverage a sequential latent variable model for knowledge selection, which subsequently improves knowledge-grounded chit-chat.

Please contact Byeongchang Kim if you have any question.

Reference

If you use this code or dataset as part of any published research, please refer following paper,

@inproceedings{Kim:2020:ICLR,
    title="{Sequential Latent Knowledge Selection for Knowledge-Grounded Dialogue}",
    author={Kim, Byeongchang and Ahn, Jaewoo and Kim, Gunhee},
    booktitle={ICLR},
    year=2020
}

System Requirements

• Python 3.6
• TensorFlow 2.0
• CUDA 10.0 supported GPU with at least 12GB memory
• see requirements.yml for more details

Running Experiments

Wizard-of-Wikipedia

To train the model from scratch,

python train.py --cfg ymls/default.yml --gpus 0,1 SequentialKnowledgeTransformer

# To run in eager mode
python train.py --cfg ymls/default.yml --gpus 0,1 --enable_function False SequentialKnowledgeTransformer

To run our pretrained model,

(it will automatically download pretrained checkpoints, or you can manually download at here)

python inference.py --cfg ymls/default.yml --gpus 0,1 --test_mode wow SequentialKnowledgeTransformer

# Will show following results
seen
{'accuracy': 0.27305699481865287,
 'kl_loss': 0.3053756,
 'knowledge_loss': 1.7310758,
 'perplexity': 53.27382,
 'rouge1': 0.19239063597262404,
 'rouge2': 0.06829999978899365,
 'rougeL': 0.1738224486787311,
 'total_loss': 6.0118966}
unseen
{'accuracy': 0.18561958184599694,
 'kl_loss': 0.27512234,
 'knowledge_loss': 2.349341,
 'perplexity': 82.65279,
 'rouge1': 0.16114443772189488,
 'rouge2': 0.04277752138282203,
 'rougeL': 0.14518138000861658,
 'total_loss': 7.039112}

Holl-E

To train the model from scratch,

python train.py --cfg ymls/holle.yml --gpus 0,1 SequentialKnowledgeTransformer

To run our pretrained model,

(it will automatically download pretrained checkpoints, or you can manually download at here or here)

python inference.py --cfg ymls/holle.yml --gpus 0,1 --test_mode holle_1 SequentialKnowledgeTransformer

# Will show following results
{'accuracy': 0.3037037037037037,
 'accuracy_multi_responses': 0.4033670033670034,
 'kl_loss': 0.36722404,
 'knowledge_loss': 1.3605422,
 'perplexity': 51.95605,
 'perplexity_multi_responses': 29.779757,
 'rouge1': 0.294273363798098,
 'rouge1_multi_responses': 0.3620479996911834,
 'rouge2': 0.22867428360364725,
 'rouge2_multi_responses': 0.2942280954677095,
 'rougeL': 0.28614673266443935,
 'rougeL_multi_responses': 0.3524777390233543,
 'total_loss': 5.678165}

 # Or you can try it with another checkpoint
python inference.py --cfg ymls/holle.yml --gpus 0,1 --test_mode holle_2 SequentialKnowledgeTransformer

Interactive Demo

You can have a chat with our SKT agent using following command (trained on Wizard-of-Wikipedia dataset),

python interactive.py --cfg ymls/default.yml --gpus 0 --test_mode wow SequentialKnowledgeTransformer

Acknowledgement

We thank Hyunwoo Kim, Chris Dongjoo Kim, Soochan Lee, and Junsoo Ha for their helpful comments.

This work was supported by SK T-Brain corporation and Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-01082, SW StarLab).

License

See LICENSE.md.