Re-DocRED-CF

Many datasets have been developed to train and evaluate document-level relation extraction (RE) models. Most of these are constructed using real-world data. However, it has been shown that RE models trained on real-world data suffer from factual biases. To evaluate and address this issue, we present CovEReD (Paper), a counterfactual data generation approach for document-level relation extraction datasets through entity replacement.

Using our pipeline, we have generated Re-DocRED-CF, a dataset of counterfactual RE documents, to help evaluate and address inconsistencies in document-level RE. This repo contains five counterfactual variations of the seed dataset, i.e., Re-DocRED. All five sets of train/dev/test dataset files are available here through the HuggingFace Datasets API 🤗.

To select a specific variation (e.g. var-01):

dataset = load_dataset("amodaresi/Re-DocRED-CF", "var-01")

Output:

DatasetDict({
    train: Dataset({
        features: ['title', 'labels', 'original_doc_id', 'vertexSet', 'sents'],
        num_rows: 2870
    })
    dev: Dataset({
        features: ['title', 'labels', 'original_doc_id', 'vertexSet', 'sents'],
        num_rows: 466
    })
    test: Dataset({
        features: ['title', 'labels', 'original_doc_id', 'vertexSet', 'sents'],
        num_rows: 453
    })
    train_mix: Dataset({
        features: ['title', 'labels', 'original_doc_id', 'vertexSet', 'sents'],
        num_rows: 5923
    })
})

The train_mix is the original training set combined with its counterfactual variation counterpart. We have also included four additional training set variations (var-[06, 07, 08, 09]), though they were not used in the evaluations presented in our paper.

The properties title, labels, vertexSet, and sents are structured similarly to those in the original DocRED & Re-DocRED datasets:

• title: Document title.
• labels: List of relations. Each entry indicates the relation between a head and a tail entity, with some entries also specifying evidence sentences.
• vertexSet: List of entity vertex sets. Each entry represents a vertex specifying all mentions of an entity by their position in the document, along with their type.
• sents: Tokenized sentences.

In examples that are counterfactually generated, the title includes a variation number. For example: AirAsia Zest ### 1. The original_doc_id denotes the index of the example in the original seed dataset, i.e., Re-DocRED.

GitHub Repo & Paper

For more information about the CovEReD pipeline, refer to:

• 📄 Paper: "Consistent Document-Level Relation Extraction via Counterfactuals"
• 🔗 GitHub Repo: https://github.com/amodaresi/CovEReD

Cite

If you use the dataset, CovEReD pipeline, or code from this repository, please cite the paper:

@inproceedings{modarressi-covered-2024,
    title="Consistent Document-Level Relation Extraction via Counterfactuals", 
    author="Ali Modarressi and Abdullatif Köksal and Hinrich Schütze",
    year="2024",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2024",
    address = "Miami, United States",
    publisher = "Association for Computational Linguistics",
}