LELNER: A Lightweight and Effective Low-resource Named Entity Recognition model

Named entity recognition aims to find the target entity from the input sentence and determine the category it belongs to. Low-resource means that the training data used by the model is scarce. In order to improve the performance of the model with a small quantity of labeled data, previous works propose the concept of the trigger and introduce trigger information. Despite their excellent performance, these methods still have shortcomings in trigger representation generation, information fusion and model training. To remedy these deficiencies, this paper proposes the LELNER model including information interaction module and information fusion network. The information interaction module realizes the interaction between trigger and sentence, which leads to trigger representation containing more entity information. The information fusion network perfectly merges the trigger representation into the sentence sequence. As a result, the network has a better fusion effect than previous nonlinear fusion methods. In terms of model training, this paper designs a one-step training method to replace the two-step training method used in previous models, which provides convenience for the entire training process. Experimental results show that the proposed LELNER model achieves state-of-the-art results on three public datasets BC5CDR, CONLL and SemEval-lap.1