CLINER: exploring task-relevant features and label semantic for few-shot named entity recognition

Few-shot named entity recognition aims at recognizing novel-class named entities in low resources scenarios. Low resource scenarios contain limited data in the support set with sparse labels. Existing methods neglect the relevance of the support set to the task and the semantics of label naming. In this paper, on the basis of contrastive learning, we propose a multi-task learning framework CLINER for Few-Shot NER. We construct a mechanism for joint learning of label semantic information and support set information. For label support set information, we find a view in the support set that is most relevant to the current task, maximizing the utilization of each support set. Momentum encoder, a dynamic queue, is constructed to keep track of positive and negative examples learned from previous support sets, and keep it updated. For label semantic information, it is implied in the label naming and is derived explicitly by pre-trained language encoder. Experiments demonstrate that our model improves the overall performance comparing with recent baseline models, achieves state-of-the-art results on the commonly used standard datasets. The source code of CLINER will be available at: https://github.com/yizumi426/CLINER.