Research On Named Entity Recognition Based On Deep Learning

Named entity recognition is one of the fundamental tasks in the field of natural language processing,which is a prerequisite for the implementation of downstream tasks such as information extraction,machine translation,intelligent question and answer,and semantic analysis,and therefore occupies an important position in the process of natural language processing technology towards full-scale application.However,the current deep learning-based named entity recognition methods mainly focus on the recognition of non-nested named entities,ignoring the semantic relationships and structural information contained in the nested named entities,so how to accurately identify the nested named entities needs to be further researched.Meanwhile,the research on non-nested named entity recognition primarily focuses on using word granularity information to improve the recognition effect,but there are problems such as over-reliance on lexicons,error transfer from a single word division tool,which leads to inefficient method domain portability and bad recognition effect,etc.Therefore,it is still a challenging task to develop algorithms that can efficiently and accurately recognize named entities.In this paper,we focus on the named entity recognition methods in general and network security domains,and divide named entities into nested and non-nested named entities according to their different structural characteristics,and propose named entity recognition algorithms with high recognition accuracy and strong domain portability from two levels of hierarchical region exhaustion and attention mechanism,The main research contents are as follows:(1)In order to solve the problem that the regular named entity recognition methods cannot recognize named entities with nested structures,which leads to the lack of fine-grained semantic and structural information in the text,which leads to the low overall accuracy of named entity recognition,we propose a layered region exhaustive Chinese nested named entity recognition method.The method uses multilayer convolutional neural networks to recursively encode the region representation of candidate entities to overcome the multi-label classification problem,and decodes the predicted labels in a hierarchical manner to alleviate layer disorientation and error propagation,as well as to improve the recognition performance of the method.Due to the lack of widely recognized datasets for Chinese nested named entity recognition,a new Chinese nested named entity recognition dataset NEPD is constructed by combining automatic generation and manual labeling methods.experimental results on this dataset show that the layered region exhaustive nested named entity recognition method can capture the association between internal and external information of nested named entities which effectively mitigates the impact of errors propagation between layers.(2)In order to solve the problems of over-reliance on lexicon and neglecting the error transfer generated by a single disambiguation tool in the non-nested named entity recognition method incorporating word granularity information,we proposed a named entity recognition method incorporating multi-source disambiguation information.This method alleviates the separation errors by aggregating multi-source separation results,encodes the separation features using a word alignment attention mechanism,and avoids the information reduction by fusing the separation features through B,M,E,and S tagging classification,thus guaranteeing the effective named entity recognition using word granularity information without relying on lexicons.Experiments on public datasets show that it has a good ability to capture word granularity information compared with similar methods,and can effectively alleviate the impact of separation errors on the recognition effect of named entities.In addition,in order to verify that the proposed method can effectively solve the problem of lack of lexicons in professional fields,which cannot effectively use word granularity information to improve recognition,the research is performed for the cyber security field.The experimental results show that the proposed method has better cybersecurity entity recognition performance compared with the basic algorithm in terms of accuracy,recall,and F1 value indexes.