Remotely supervised named entity recognition (DS-NER) is widely used in real-world scenarios, but it suffers from the problem of label noise. Existing methods based on the teacher-student framework have limitations in that they generate incorrect pseudo-labeled samples due to the low reliability of the teacher network, which leads to error propagation. To address these issues, in this paper, we propose (1) uncertainty-aware teacher learning to reduce the number of incorrect pseudo-labels by exploiting prediction uncertainty, and (2) student-student collaborative learning to reduce the reliance on pseudo-labels and fully explore the mislabeled samples through reliable label propagation between the two student networks. The proposed method outperforms the state-of-the-art DS-NER methods on five DS-NER datasets.
