Incorporating boundary and category feature for nested named entity recognition

In the natural language processing (NLP) field, it is fairly common that an entity is nested in another entity. Most existing named entity recognition (NER) models focus on flat entities but ignore nested entities. In this paper, we propose a neural model for nested named entity recognition. Our model employs a multi-label boundary detection module to detect entity boundaries, avoiding boundary detection conflict existing in the boundary-aware model. Besides, our model with a boundary detection module and a category detection module detects entity boundaries and entity categories simultaneously, avoiding the error propagation problem existing in current pipeline models. Furthermore, we introduce multitask learning to train the boundary detection module and the category detection module to capture the underlying association between entity boundary information and entity category information. In this way, our model achieves better performance of entity extraction. In evaluations on two nested NER datasets and a flat NER dataset, we show that our model outperforms previous state-of-the-art models on nested and flat NER.