NoisyNER

NoisyNER is a dataset for the evaluation of methods to handle noisy labels when training machine learning models. It is from the NLP/Information Extraction domain and was created through a realistic distant supervision technique. Some highlights and interesting aspects of the data are: - Seven sets of labels with differing noise patterns to evaluate different noise levels on the same instances - Full parallel clean labels available to compute upper performance bounds or study scenarios where a small amount of gold-standard data can be leveraged - Skewed label distribution (typical for Named Entity Recognition tasks) - For some label sets: noise level higher than the true label probability - Sequential dependencies between the labels For more details on the dataset and its creation process, please refer to our publication ojs.aaai.org/index.php/AAAI/article/view/16938 (published at AAAI'21).