NeuS: Neutral Multi-News Summarization for Mitigating Framing Bias

Code for "NeuS: Neutral Multi-News Summarization for Mitigating Framing Bias", NAACL2022 [PDF]

If you use any source codes or datasets included in this toolkit in your work, please cite the following paper. The bibtex is listed below:

@inproceedings{lee2022neus,
  title={NeuS: Neutral Multi-News Summarization for Mitigating Framing Bias},
  author={Lee, Nayeon and Bang, Yejin and Yu, Tiezheng and Madotto, Andrea and Fung, Pascale},
  journal={Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL)},
  year={2022}
}

0. Setup

Run pip install -r requirement.txt

Acknowledgement: Our codes are based on Huggingface's Transformer codebase.

1. Dataset & Pre-processing

How?

Run data_preprocessing.ipynb.

• To replicate all our data preprocessing steps, run all codes.
• To just prepare testset of your own to run with our pretrained model, run the create_source_target_with_processing_probe_format(article_test, 'test') function with your own article_test list.

Additional Details:

Datasets are inside data folder:

• raw_crawled/ : contains crawled data from Allsides.com until 2021-10-19 (with all the meta dataset)
• naacl2022_filtered_allsides_article.json : filtered & preprocessed verion from raw_crawled.
• naacl2022_lrc_roundup_random_order_probe : contains final train/val/test files used in our NeuS-Title model.

Full article version (smaller subset): Not directly used in our paper, but releasing to help the community 😊

• BASIL (cite) extended (neutral): Refer to https://github.com/nayeon7lee/framing_bias_datasets
• AllSides articles: Refer to https://github.com/nayeon7lee/framing_bias_datasets

2. Train Model

Run bash finetune_neus_script.sh with desired hyper-parameters & correct configurations.

3. Generate neutral summary

1. Download our checkpoint from here into save/ directory
2. Run bash generate_script.sh with correct configurations (i.e. PROJECT_DIR, OUTPUT_DIR, DATA_DIR).

If you want to generate using your own model checkpoint, change the OUTPUT_DIR to be the directory containing your model checkpoint.

4. Evaluate (Metric)

Lexicon Resources:

All the lexicons are inside data/lexicons:

• NRC-VAD-Lexicon-Aug2018Release/ : VAD lexicon used in our paper. Refer to the README inside the folder for citation details.
• bias-lexicon/ : other lexicons that the community could find useful (not used in our work). Refer to README inside the folder for citation details.
• bias_related_lexicons/ : other lexicons that the community could find useful (not used in our work). Refer to README inside the folder for citation details.

Running the evaluation:

Run the codes in metric_code.ipynb.

• Make sure you have data/lexicons/NRC-VAD-Lexicon-Aug2018Release/.
• Also, make sure you correctly load test generation texts and target texts into gens and test_tgts variables.