Корсар

AbstractToday’s text classifiers inevitably suffer from unintended dataset biases, especially the document-level label bias and word-level keyword bias, which may hurt models’ generalization. Many previous studies employed data-level manipulations or …...