whl97 / LS-Score

This project is the implementation of paper Unsupervised Reference-Free Summary Quality Evaluation via Contrastive Learning. (Paper Link)

Datasets

Newsroom

the complete dataset is available here.

The processed dataset is available here.

CNN/Daily Mail

the complete dataset is also available here.

The processed dataset is available here.

Fine-tuning

1. Download the pretrained model and modified pretrained_model_dir. We use bert-base-uncased as our base model to finetune.

2. Download the datasets and modified dataset_dir_dict.

3. To fine-tune the model:

   python3 code/trainer.py --dataset_name=cnndm
   python3 code/trainer.py --dataset_name=newsroom