metadata

license: apache-2.0

AVeriTeC

Data, knowledge store and source code to reproduce the baseline experiments for the AVeriTeC dataset, which will be used for the 7th FEVER workshop co-hosted at EMNLP 2024.

Dataset

The training and dev dataset can be found under data. Test data will be released at a later date. Each claim follows the following structure:

{
    "claim": "The claim text itself",
    "required_reannotation": "True or False. Denotes that the claim received a second round of QG-QA and quality control annotation.",
    "label": "The annotated verdict for the claim",
    "justification": "A textual justification explaining how the verdict was reached from the question-answer pairs.",
    "claim_date": "Our best estimate for the date the claim first appeared",
    "speaker": "The person or organization that made the claim, e.g. Barrack Obama, The Onion.",
    "original_claim_url": "If the claim first appeared on the internet, a url to the original location",
    "cached_original_claim_url": "Where possible, an archive.org link to the original claim url",
    "fact_checking_article": "The fact-checking article we extracted the claim from",
    "reporting_source": "The website or organization that first published the claim, e.g. Facebook, CNN.",
    "location_ISO_code": "The location most relevant for the claim. Highly useful for search.",
    "claim_types": [
            "The types of the claim",
    ],
    "fact_checking_strategies": [
        "The strategies employed in the fact-checking article",
    ],
    "questions": [
        {
            "question": "A fact-checking question for the claim",
            "answers": [
                {
                    "answer": "The answer to the question",
                    "answer_type": "Whether the answer was abstractive, extractive, boolean, or unanswerable",
                    "source_url": "The source url for the answer",
                    "cached_source_url": "An archive.org link for the source url"
                    "source_medium": "The medium the answer appeared in, e.g. web text, a pdf, or an image.",
                }
            ]
        },
    ]
}

Reproduce the baseline

Below are the steps to reproduce the baseline results. The main difference from the reported results in the paper is that, instead of requiring direct access to the paid Google Search API, we provide such search results for up to 1000 URLs per claim using different queries, and the scraped text as a knowledge store for retrieval for each claim. This is aimed at reducing the overhead cost of participating in the Shared Task. Another difference is that we also added text scraped from pdf URLs to the knowledge store.

0. Set up environment

You will need to have Git LFS installed:

git lfs install
git clone https://huggingface.co/chenxwh/AVeriTeC

You can also skip the large files in the repo and selectively download them later:

GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/chenxwh/AVeriTeC

Then create conda environment and install the libs.

conda create -n averitec python=3.11
conda activate averitec

pip install -r requirements.txt
python -m spacy download en_core_web_lg
python -m nltk.downloader punkt
python -m nltk.downloader wordnet
conda install pytorch pytorch-cuda=11.8 -c pytorch -c nvidia

1. Scrape text from the URLs obtained by searching queries with the Google API

The URLs of the search results and queries used for each claim can be found here.

Next, we scrape the text from the URLs and parse the text to sentences. The processed files are also provided and can be found here. You can use your own scraping tool to extract sentences from the URLs.

bash script/scraper.sh <split> <start_idx> <end_idx> 
# e.g., bash script/scraper.sh dev 0 500

2. Rank the sentences in the knowledge store with BM25

Then, we rank the scraped sentences for each claim using BM25 (based on the similarity to the claim), keeping the top 100 sentences per claim. See bm25_sentences.py for more argument options. We provide the output file for this step on the dev set here.

python -m src.reranking.bm25_sentences

3. Generate questions-answer pair for the top sentences

We use BLOOM to generate QA paris for each of the top 100 sentence, providing 10 closest claim-QA-pairs from the training set as in-context examples. See question_generation_top_sentences.py for more argument options. We provide the output file for this step on the dev set here.

python -m src.reranking.question_generation_top_sentences

4. Rerank the QA pairs

Using a pre-trained BERT model bert_dual_encoder.ckpt, we rerank the QA paris and keep top 3 QA paris as evidence. See rerank_questions.py for more argument options. We provide the output file for this step on the dev set here.

python -m src.reranking.rerank_questions

5. Veracity prediction

Finally, given a claim and its 3 QA pairs as evidence, we use another pre-trained BERT model bert_veracity.ckpt to predict the veracity label. See veracity_prediction.py for more argument options. We provide the prediction file for this step on the dev set here.

python -m src.prediction.veracity_prediction

Then evaluate the veracity prediction performance with (see evaluate_veracity.py for more argument options):

python -m src.prediction.evaluate_veracity

The result for dev and the test set below. We recommend using 0.25 as cut-off score for evaluating the relevance of the evidence.

Model	Split	Q only	Q + A	Veracity @ 0.2	@ 0.25	@ 0.3
AVeriTeC-BLOOM-7b	dev	0.24	0.19	0.19	0.09	0.05
AVeriTeC-BLOOM-7b	test

Citation

If you find AVeriTeC useful for your research and applications, please cite us using this BibTeX:

@article{schlichtkrull2024averitec,
  title={Averitec: A dataset for real-world claim verification with evidence from the web},
  author={Schlichtkrull, Michael and Guo, Zhijiang and Vlachos, Andreas},
  journal={Advances in Neural Information Processing Systems},
  volume={36},
  year={2024}
}