Multilayer ToI Detection Approach for Nested NER

Nested entities commonly exist in news articles and biomedical corpora. The performance of nested NER is still a great challenge in the field of named entity recognition (NER). Unlike the structural models in previous work, this paper presents a comprehensive study of nested NER by means of text-of-interest (ToI) detection. This paper presents a novel ToI-CNN with dual transformer encoders (ToI-CNN + DTE) model for this solution. We design a directional self-attention mechanism to encode contextual representation over the whole-sentence in the forward and backward directions. The features of the entities are extracted from the contextual token representations by a convolutional neural network. Moreover, we use HAT pooling operation to convert the various length ToIs to a fixed length vector and connect with a fully connected network for classification. The layer where the nested entities are located can be evaluated by multi-task learning jointly with layer classification. The experimental results show that our model achieves excellent performance in F1 score, training cost and layer evaluation on the nested NER datasets.