Triaffine Decoder for span-based Chinese Named Entity Recognition

Abstract In recent years, some advanced NER (Named Entity Recognition) models have achieved satisfactory results, but modern neural networks are often plagued by the over-confidence problem. Driven by the lack of entity information in the span-based NER model, we perform lexical enhancement operations for Chinese in the decoding layer of the span-based NER model. Roughly speaking, our model builds on the advanced span-based NER model by utilizing the Triaffine structure to construct a tree-shaped scoring mechanism for the decoding process of entities, and at the same time adding a softlexicon embedding in the input layer of the model. The improved model can introduce more entity information, obtain more advanced performance than before, and alleviate the over-confidence problem. Our model experiments on four well-known Chinese NER datasets and the F1 values on three of them achieve competitive results compared to the baseline model.