SciFact

This repository contains data and code for the paper Fact or Fiction: Verifying Scientific Claims by David Wadden, Shanchuan Lin, Kyle Lo, Lucy Lu Wang, Madeleine van Zuylen, Arman Cohan, and Hannaneh Hajishirzi.

⬇️Download the dataset here.

🏆Participate in SCIVER shared task here.

📈Check out the leaderboard here.

You can also check out our COVID-19 claim verification demo. For a heavier-weight COVID claim verifier, see the section on verifying COVID-19 claims.

Update (May 2022): The MultiVerS model, which achieves SOTA on SciFact and two other scientific claim verification datasets, is available at https://github.com/dwadden/multivers.

Update (Dec 2020): SciFact will be used for the SciVer shared task to be featured at the SDP workshop at NAACL 2021. Registration is open!

Update (Dec 2020): Claim / citation context data now available to train claim generation models. See Claim generation data.

Table of contents

• Leaderboard
• Dependencies
• Run models for paper metrics
• Dataset
• Download pre-trained models
• Training scripts
• Verify claims about COVID-19
• Citation
• Contact

Leaderboard

UPDATE (Jan 2021): We now have an official AI2 leaderboard with automated evaluation! For information on the submission file format and evaluation metrics, see evaluation.md. Or, check out the getting started page on the leaderboard.

Dependencies

We recommend you create an anaconda environment:

conda create --name scifact python=3.7 conda-build

Then, from the scifact project root, run

conda develop .

which will add the scifact code to your PYTHONPATH.

Then, install Python requirements:

pip install -r requirements.txt

Run models for paper metrics

We provide scripts let you easily run our models and re-create the dev set metrics published in paper. The script will automatically download the dataset and pre-trained models. You should be able to reproduce our dev set results from the paper by following these instructions (we are not releasing test set labels at this point). Please post an issue if you're unable to do this.

To recreate Table 3 rationale selection metrics:

./script/rationale-selection.sh [bert-variant] [training-dataset] [dataset]

To recreate Table 3 label prediction metrics:

./script/label-prediction.sh [bert-variant] [training-dataset] [dataset]

• [bert-variant] options: roberta_large, roberta_base, scibert, biomed_roberta_base
• [training-dataset] options: scifact, scifact_only_claim, scifact_only_rationale, fever_scifact, fever, snopes
• [dataset] options: dev.

To make full-pipeline predictions, you can use:

./script/pipeline.sh [retrieval] [model] [dataset]

• [retrieval] options: oracle, open
• [model] options: oracle-rationale, zero-shot, verisci
• [dataset] options: dev, test.

Two notes on this:

• For the dev set, this script will also compute performance metrics. For the test set the "gold" labels are not public, so the script will just make predictions without evaluating.
• oracle retrieval will break on the test set, since it requires access to the gold evidence documents. But open retrieval will work on both dev and test.

Make full-pipeline predictions

Since the test set is not public, you can't recreate the test set metrics reported in Table 4. However, you can make predictions on the test set; see above.

Dataset

Download with script: The data will be downloaded and stored in the data directory.

./script/download-data.sh

Or, click here to download the tarball.

The claims are split into claims_train.jsonl, claims_dev.jsonl, and claims_test.jsonl, one claim per line. The claim and dev sets contain labels, while the test set is unlabeled. For test set evaluation, submit to the leaderboard! The corpus of evidence documents is corpus.jsonl, one evidence document per line.

Due to the relatively small size of the dataset, we also provide a 5-fold cross-validation split that may be useful for model development. After unzipping the tarball, the data will organized like this:

data
| corpus.jsonl
| claims_train.jsonl
| claims_dev.jsonl
| claims_test.jsonl
| cross_validation
  | fold_1
    | claims_train_1.jsonl
    | claims_dev_1.jsonl
  ...
  | fold_5
    | claims_train_5.jsonl
    | claims_dev_5.jsonl

See data.md for descriptions of the schemas for each file type.

Claim generation data

We also make available the collection of claims together with the documents and citation contexts they are based on. We hope that these data will facilitate the training of "claim generation" models that can summarize a citation context into atomic claims. Click here to download the file, or enter

wget https://scifact.s3-us-west-2.amazonaws.com/release/latest/claims_with_citances.jsonl -P data

For more information on the data, see claims-with-citances.md

Download pre-trained models

All "BERT-to-BERT"-style models as described in the paper are stored in a public AWS S3 bucket. You can download the models using the script:

./script/download-model.sh [model-component] [bert-variant] [training-dataset]

• [model-component] options: rationale, label
• [bert-variant] options: roberta_large, roberta_base, scibert, biomed_roberta_base
• [training-dataset] options: scifact, scifact_only_claim, scifact_only_rationale, fever_scifact, fever, snopes

The script checks to make sure the downloaded model doesn't already exist before starting new downloads.

The best-performing pipeline reported in paper uses:

• rationale: roberta_large + scifact
• label: roberta_large + fever_scifact

For fever and fever_scifact, there are models available for all 4 BERT variants. For snopes, only roberta_large is available for download (but you can train your own model).

After downloading the pretrained-model, you can follow instruction model.md to run individual model components.

Training scripts

See training.md.

Verify claims about COVID-19

While the project website features a COVID-19 fact-checking demo, it is not configurable and uses a "light-weight" version of VeriSci based on DistilBERT. We provide a more configurable fact-checking script that uses the full model. Like the web demo, it uses covidex for document retrieval. Usage is as follows:

python script/verify_covid.py [claim-text] [report-file] [optional-arguments].

For a description of the optional arguments, run python script/verify_covid.py -h. The script generates either a pdf or markdown report. The pdf version requires pandoc and wkhtmltopdf, both of which can be installed with conda. A usage example might be:

python script/verify_covid.py \
  "Coronavirus droplets can remain airborne for hours" \
  results/covid-report \
  --n_documents=100 \
  --rationale_selection_method=threshold \
  --rationale_threshold=0.2 \
  --verbose \
  --full_abstract

The 36 claims COVID claims mentions in the paper can be found at covid/claims.txt.

Citation

@inproceedings{Wadden2020FactOF,
  title={Fact or Fiction: Verifying Scientific Claims},
  author={David Wadden and Shanchuan Lin and Kyle Lo and Lucy Lu Wang and Madeleine van Zuylen and Arman Cohan and Hannaneh Hajishirzi},
  booktitle={EMNLP},
  year={2020},
}

Contact

Email: davidw@allenai.org.