Research And Application Of Regression Network Method For Nested Named Entity Recognitio

Named entity recognition is one of the basic tasks in natural language processing.In natural language texts,there are a large number of named entities with nested structures.Traditional methods using sequence annotation models for entity recognition cannot effectively identify named entities with nested structures.In order to effectively identify nested named entities,this paper designs a regression network model that can end-to-end identify entities with nested structures.The research work of this article can be divided into the following two parts for introduction:(1)Propose a nested named entity recognition model based on regression network(RN).The model first maps sentences to a saturated higher-order abstract semantic representation through a deep neural network.Then generate candidate text bounding boxes on this abstract semantic representation.Finally,drawing inspiration from the idea of object detection algorithms in the image field,a text based regression algorithm was designed to construct an end-to-end multi-objective learning framework.This framework can simultaneously predict the classification scores of entity candidates and refine their spatial positions in sentences,and can optimize the model globally by sharing parameters.Through experiments,it has been shown that the RN model outperforms the relevant work by about 5% and 2% in F1 values on the ACE2005 Chinese and English datasets,respectively,and has also achieved competitive performance on the GENIA dataset.(2)Propose a full length nested named entity recognition model based on regression networks.The previous model required enumerating a large number of text borders,resulting in high computational complexity and inability to identify long entities through regression.However,there are a large number of long entities in real text.If the method of the previous model is used to identify long entities,the semantic weakening is severe,and generating candidate entities can easily lead to high computational complexity and data imbalance.To address these two issues,this paper proposes a multi granularity full length nested named entity recognition model based on regression networks,which does not need to exhaustively list all possible candidate entity spans,Only a small portion of the span needs to be verified and all possible entities in the sentence can be identified through regression operations.In addition,for long entities,multi granularity sentence representations are used to encode semantic features of different semantic granularity.In this experiment,even if a small number of entity spans were enumerated,the model still remained competitive,achieving 87.35% and 80.85% F1 scores on the ACE2005 and GENIA datasets.