Nested Named Entity Recognition Based On Controlled Attention Metho

Named entity recognition is one of the basic tasks in natural language processing.Due to the characteristics of natural language,there are a large number of nested named entities in text.Traditional methods use sequential annotation to identify named entities,but can not identify entities with nested structure.For the recognition of nested named entities,some models based on span and hypergraph are proposed.These methods can solve the problem of nested entity recognition,but they can not fully perceive the entity boundary and encode the semantic dependency characteristics of entity boundary and context.Based on the idea of controlled attention in cognitive neuroscience,a nested named entity recognition method based on controlled attention is proposed in this paper.The research work can be divided into the following two parts:(1)A nested named entity recognition method based on controlled attention labels is proposed.In cognitive neuroscience,controlled attention is defined as the ability to direct attention to a specific location in the stimulus field.Relevant studies have shown that focusing on one stimulus can improve the efficiency of processing targets.Inspired by controlled attention,this paper uses controlled attention labels to distinguish different named entities.The method first generates controlled attention labels(entity boundary information)and candidate spans.Then,the controlled attention labels is embedded in the original sentence to identify the position of the candidate spans.Finally,the sentences with controlled attention labels are classified.(2)A nested named entity recognition method based on controlled attention labels regression is proposed.This method uses regression operation to predict the position offset of candidate entities relative to real entities,and can dynamically adjust and control the position of attention labels in sentences during training.The related research of cognitive neuroscience also shows that the cognitive process is very complicated,and locating the stimulus source is a step-by-step process.Based on the controlled attention labels model,this paper proposes a model based on controlled attention labels regression,and adopts the method of iterative regression,that is,after updating the position of the controlled attention labels,the regression continues until the end of the iteration.Controlled attention labels can not only solve the problems of nested entity representation and loss of global semantic relationship,but also encode the semantic structure dependency information of candidate spans and context.In addition,for the candidate spans with high coincidence with the real entity,iterative regression controlled attention labels can effectively dynamically adjust the candidate spans to the position of the real entity,and further simulate the human cognitive process.Through verification on public data sets,the nested named entity recognition method based on controlled attention has obvious performance advantages on both Chinese and English datasets.