Mind the Gap! Injecting Commonsense Knowledge for Abstractive Dialogue Summarization

The official repository for the paper "Mind the Gap! Injecting Commonsense Knowledge for Abstractive Dialogue Summarization" accepted at COLING 2022.

Paper Link : https://arxiv.org/abs/2209.00930

Youtube Explanation : https://www.youtube.com/watch?v=xHr3Ujlib4Y

Overview of method, SICK (Summarizing with Injected Commonsense Knowledge).

Setting

The following command will clone the project:

git clone https://github.com/SeungoneKim/SICK_Summarization.git

Before experimenting, you can make a virtual environment for the project.

conda create -n sick python=3.8
conda activate sick
pip install torch==1.10.1+cu113 torchvision==0.11.2+cu113 torchaudio==0.10.1 -f https://download.pytorch org/whl/cu113/torch_stable.html
pip install -r requirements.txt
pip install -U spacy
python -m spacy download en_core_web_sm

Dataset Download

For training and evaluating on Samsum, we use dataset provided by Hugging Face Dataset Library. For Dialogsum, the dataset is not automatically provided, so you can download it from the url below,

https://drive.google.com/drive/folders/1CuZaU5Xw0AiIPaBTRrjToFkktS7_6KwG?usp=share_link

and put it under the directory of SICK_summarization/data/DialogSum_Data.

mkdir data/DialogSum_Data

Also, you could download the preprocessed commonsense data from the url below,

https://drive.google.com/drive/folders/1z1MXBGJ3pt0lC5dneMfFrQgxXTD8Iqrr?usp=share_link

and put it under the directory of SICK_summarization/data/COMET_data.

mkdir data/COMET_data

To process the commonsense data, we used COMET-ATOMIC 2020 and PARACOMET github repository. Huge thanks to the authors for providing the awesome code:)

Training & Evaluation

You can use the following commands to train either SICK, SICK++. Note that for SICK++, we use the customized model class, BartForConditionalGeneration_DualDecoder and the customized trainer class DualDecoderTrainer each located in models/bart.py and src/trainer.py. It will automatically loaded if you set the following arguments correctly.

For training SICK, you can use the following command.

CUDA_VISIBLE_DEVICES="1" python3 train_summarization_context.py --finetune_weight_path="./new_weights_sick" --best_finetune_weight_path="./new_weights_sick_best" --dataset_name="samsum" --use_paracomet=True --model_name="facebook/bart-large-xsum" --relation "xIntent" --epoch=1 --use_sentence_transformer True

Some configs to specify.

• dataset_name : Specify either "samsum" or "dialogsum"
• use_paracomet : If you set to true, it will use paracomet, and if not, it will use comet by default.
• use_sentence_transformer : If you would like to use the commonsense selected with sentence_transformer, you should use this argument
• use_roberta : If you would like to use the commonsense selected with roberta_nli model, you should use this argument.
• relation : If you would only like to use one of the 5 possible relations, you could specify it with this argument.

For training SICK++, you can use the following command.

CUDA_VISIBLE_DEVICES="1" python3 train_summarization_full.py --finetune_weight_path="./new_weights_sickplus" --best_finetune_weight_path="./new_weights_sickplus_best" --dataset_name="samsum" --use_paracomet=True --model_name="facebook/bart-large-xsum" --relation "xIntent" --supervision_relation "xIntent" --epoch=1 --use_sentence_transformer True

Some configs to specify.

• dataset_name : Specify either "samsum" or "dialogsum"
• use_paracomet : If you set to true, it will use paracomet, and if not, it will use comet by default.
• use_sentence_transformer : If you would like to use the commonsense selected with sentence_transformer, you should use this argument
• use_roberta : If you would like to use the commonsense selected with roberta_nli model, you should use this argument.
• relation : If you would only like to use one of the 5 possible relations, you could specify it with this argument.
• supervision_relation : If you would only like to use one of the 5 possible supervision relations, you could specify it with this argument.

For inferencing either SICK or SICK++, you could use the following command.

CUDA_VISIBLE_DEVICES="1" python3 inference.py --dataset_name "samsum" --model_checkpoint="./new_weights_paracomet_best" --test_output_file_name="./tmp_result.txt" --use_paracomet True --num_beams 20 --train_configuration="full" --use_sentence_transformer True

Some configs to specify.

• dataset_name : Specify either "samsum" or "dialogsum"
• use_paracomet : If you set to true, it will use paracomet, and if not, it will use comet by default.
• use_sentence_transformer : If you would like to use the commonsense selected with sentence_transformer, you should use this argument
• use_roberta : If you would like to use the commonsense selected with roberta_nli model, you should use this argument.
• relation : If you would only like to use one of the 5 possible relations, you could specify it with this argument.
• supervision_relation : If you would only like to use one of the 5 possible supervision relations, you could specify it with this argument.
• num_beams : beam size during decoding
• model_checkpoint : The checkpoint to use during inference(test) time. We suggest to use the best_finetune_weight_path during training.
• training_configuration : If you want to test SICK, type "context". If you want to test SICK++, type "full".

Citation

If you find this useful, please consider citing our paper:

@inproceedings{kim2022mind,
  title={Mind the Gap! Injecting Commonsense Knowledge for Abstractive Dialogue Summarization},
  author={Kim, Seungone and Joo, Se June and Chae, Hyungjoo and Kim, Chaehyeong and Hwang, Seung-won and Yeo, Jinyoung},
  booktitle={Proceedings of the 29th International Conference on Computational Linguistics},
  pages={6285--6300},
  year={2022}
}

Point of contact

For any questions about the implementation or content of the paper, you could contact me via the following email:)

louisdebroglie@kaist.ac.kr