| Named Entity Recognition(NER)is a crucial task in natural language processing that has been extensively studied in both academic and industrial communities.Despite the significant success achieved by existing deep learning-based NER methods,they have certain limitations in capturing semantic associations,which can result in incomplete feature extraction.Specifically,two limitations have been identified.Firstly,existing methods often neglect fine-grained information,such as affix semantics.Traditional methods typically extract features at the character or word level,disregarding the clear semantic information conveyed by affixes in English words.This can hinder the effective capture of internal semantic features of words when relying solely on character embeddings.Secondly,there is a lack of emphasis on boundary-related semantics,where the identification of start and end boundaries is treated as two isolated tasks,without sufficient attention to their relationship.To address these issues,this paper proposes NER methods that integrate affixrelated semantics and boundary-related semantics through in-depth research from these two perspectives.Specifically,the NER method based on affix-related semantics utilizes an attention mechanism to integrate affix information with the text information,enabling the method to focus on the internal information of words and supplement their semantics.The effectiveness of this method has been demonstrated on both general and biological datasets.On the other hand,the NER method based on boundaryrelated semantics leverages BERT as the encoding layer and focuses on boundary recognition tasks,where the entity end boundary recognition is performed based on the identification of entity start boundaries,enabling the method to capture boundaryrelated information.Experimental results indicate that this method significantly improves the performance of the NER task.In conclusion,this paper contributes to the field of NER by proposing novel methods that address the limitations of existing deep learning-based approaches.These methods incorporate fine-grained affix information and boundary-related semantics,which have been demonstrated to enhance the feature extraction process and improve NER performance on both general and specialized datasets.The proposed methods have the potential to facilitate various applications,such as word search and contract information extraction,and can advance the state-of-the-art in NER research. |
