jgvb99/wl-coref-gender-neutral

This Dutch coreference resolution model is based on the wl-coref model. The model was debiased through CDA, in order to improve its performance on gender-neutral pronouns and neopronouns. We used XLM-RoBERTa-base as our base model, fine-tuned the model on the SoNaR-1 corpus, and then further fine-tuned the model on a gender-neutral version of the this corpus for debiasing purposes. We used five different seeds during our experiment, and upload all five versions of the model.

For more information, see our Github repository.

This model was created as part of our FAccT 2024 paper. Find our published paper at https://dl.acm.org/doi/10.1145/3630106.3659049 and arxiv paper at https://arxiv.org/abs/2405.00134.

Usage

For usage instructions for the wl-coref model, see their repo.

Citation

G. van Boven, Y. Du, D. Nguyen, Transforming Dutch: Debiasing Dutch Coreference Resolution Systems for Non-binary Pronouns. FAccT 2024.

@inproceedings{boven-2024-transforming,
title = "Transforming Dutch: Debiasing Dutch Coreference Resolution Systems for Non-binary Pronouns",
author = "van Boven, Goya and Du, Yupei and Nguyen, Dong",
booktitle = "Proceedings of the 2024 Conference on Fairness, Accountability, and Transparency",
month = jun,
year = "2024",
address = "Online and Rio de Janeiro, Brazil",
publisher = "Association for Computing Machinery"
}