Research on NER Based on Register Migration and Multi-task Learning

Abstract

The insufficient number of tags is currently the biggest constraint on named entity recognition (NER) technology, with only a small number of Registers (means the domain of language, which will be explained in Part I) currently having a corpus with sufficient tags. The linguistic features of different Registers vary greatly, and thus a corpus with sufficient labels cannot be applied to NER in other Registers. In addition, most of the current NER models are more designed for large samples with sufficient labels, and these models do not work well in small samples with a small number of labels. To address the above problems, this paper proposes a model T_NER based on the idea of migration learning and multi-task learning, which learns the common features of language by using the idea of multi-tasking, and passes the model parameters of neurons with common features of language learned from multiple well-labelled source domains to the neurons in the target domain to achieve migration learning based on parameter sharing. In baseline experiments, T_NER’s neurons outperformed the original models such as BiLSTM and BiGRU on a small-sample NER task; in formal experiments, the more the Registers in source domains, the better T_NER’s recognition of the target domain. The experiments demonstrate that T_NER can achieve NER for small samples and across Registers.