RepEKShot: an evidential k-nearest neighbor classifier with repulsion loss for few-shot named entity recognition

Metric-based models have recently shown promising performance in the few-shot named entity recognition (NER) task. Many methods train their encoders with loss functions that focus on distinguishing different entity types, which ignores improving the ability to recognize ground-truth and interfered labels when making predictions. Furthermore, the inference strategy of nearest neighbor is popular for metric-based models. However, other surrounding neighbors can also provide useful information for NER, and it is hard to determine whether the nearest neighbor is the most suitable referent when multiple neighbors are all close to the query sample. To solve the above problems, we propose RepEKShot, a novel model which utilizes repulsion loss for training the encoder and extends the inference strategy from nearest neighbor to evidential k-nearest neighbor in the framework of Dempster–Shafer theory. Our model effectively optimizes the training of encoder, and sufficiently exploits the information provided by other neighbors to provide a more global perspective for few-shot NER. Extensive experiments have been conducted on two benchmarks with public datasets, and the results show that our model has performance merits in few-shot scenarios.