Research On End-to-End Nested Named Entity Recognition Metho

The traditional named entity recognition method usually uses the classification method to classify each word in the sentence.The semantic roles of words in named entities can be identified by classifying tags,such as the beginning and end of the entity.In order to obtain the structural features of sentences,traditional methods use sequence labeling models(e.g.,hidden Markov,conditional random fields,recurrent neural networks,etc.)to output a maximized tag sequence.The disadvantage is that there are a lot of nested entities in the text.In nested entities,a word may belong to multiple entities at the same time,with multiple labels.Output a maximized annotation path unrecognized in nested named entities.In order to effectively identify nested named entities,this paper proposes nested named entity recognition algorithm based on centrality and nested named entity recognition algorithm based on multi-scale feature learning,respectively,as follows:(1)Nested named entity recognition algorithm based on centrality: this method first maps sentences to abstract representations.Then,each word is taken as the center of the entity,and the offset of the left and right boundaries with respect to the center of the entity is calculated by regression operation.Because there is no center overlap between nested entities,this method can effectively solve the problem of nested named entity recognition and effectively suppress the generated low quality boundary.On this basis,we design an end-to-end multi-objective learning framework,which simultaneously predicts the categories of entities and their positions in sentences,and achieves global optimization by sharing parameters.Verified in ACE2005 Chinese dataset,the proposed method can effectively identify nested named entities.(2)Nested named entity recognition algorithm based on multi-scale feature learning: The previous method needs to identify the position offset of entity boundary with respect to the center point.However,in long entities,the entity boundary is far from the center,and the problem of semantic weakening is more serious.To solve this problem,we propose a multi-scale feature learning algorithm.In this method,the deep convolutional neural network is used to extract the feature from the context feature map and generate multi-scale sentence representation.Then,entities of different length are predicted in multi-scale sentence representation.This method can effectively detect named entities of different lengths.In addition,Fovea sampling algorithm is designed to replace the centrality algorithm.This method transfers the task of suppressing low-quality boundary from the training level to the sampling level,which can suppress low-quality boundary more effectively and reduce the difficulty of network fitting.Experimental results show that the improved algorithm can significantly improve the performance of recognition.