CLEANCONLL

CLEANCONLL是由柏林洪堡大学的Susanna Rücker和Alan Akbik创建的一个近无噪声的命名实体识别数据集，是对著名的CoNLL-03数据集的改进版本。该数据集通过引入实体链接信息，增强了标注的一致性和质量，使得现有的最先进模型在该数据集上能达到更高的F1分数（97.1%），并且由于标注噪声的减少，正确预测被错误计为错误的比例从47%降至6%。CLEANCONLL适用于分析最先进模型的剩余错误，并且表明高资源、粗粒度的NER任务的理论上限尚未达到。该数据集已公开发布，供研究社区使用，以促进NER领域的进一步研究和模型评估。

CLEANCONLL is a near-noise-free named entity recognition (NER) dataset developed by Susanna Rücker and Alan Akbik from Humboldt-Universität zu Berlin, which serves as an enhanced revision of the widely recognized CoNLL-03 dataset. By integrating entity linking information, this dataset improves annotation consistency and quality, enabling contemporary state-of-the-art models to achieve a superior F1 score of 97.1% on this benchmark. Furthermore, with the reduction of annotation noise, the proportion of correct predictions erroneously counted as errors has dropped from 47% to 6%. CLEANCONLL is suitable for analyzing the residual errors of state-of-the-art NER models, and it demonstrates that the theoretical upper bound of high-resource, coarse-grained NER tasks remains unachieved. This dataset has been publicly released for the global research community to facilitate further advancements in NER research and model evaluation.