Transfer Learning In Named Entity Recognition

Named entity recognition is a very basic and important task in the field of natural language processing,which aims to identify named entity words in a given text.This task can be used by some downstream tasks,such as: relationship extraction,time extraction,question and answer,etc.However,this task relies too much on a large number of labeled corpora.In the traditional field,due to sufficient corpus,the quality of corpus annotation is good.The identification of names,place names and organization names usually have better performance,but in some specific areas(such as tourism).The field,the medical field),limited by the labeled data,the effect of named entity recognition has a significant decline compared to the traditional field.How to carry out domain migration and improve the performance of specific domain models is the main research content of this topic.The data set used in this paper is from the large-scale military field corpus and small-scale traditional field corpus provided by Yunfu Technology.The military field is the source field,and the traditional field is the target field.Three different migration learning methods are used to complete the domain migration task of named entity recognition.Based on the parameter migration,the method of identifying the domain migration of the named entity is performed by performing the Finetune on the pretrained model of the source domain,and training the word vectors of the two fields separately to achieve the purpose of distinguishing the vector space.Testing on test data sets is better than traditional named entity recognition models.Based on the multi-task learning method,the domain name mapping mechanism is added to the military field and the traditional field of Yunfu Technology.Better extract specific domain characteristics.The final result is superior to the traditional named entity recognition model.Based on the domain migration model of named entity recognition against learning,considering that the part of the shared feature in the multi-task learning framework contains noise,that is,the characteristics of specific domains will be doped with other domain features,this paper joins the anti-learning mechanism and further shares the features.Purification,in addition,the pooling layer is added for feature abstraction and gradient inversion layer to optimize the discriminator parameters.The relationship between words and words in the abstract sentence of the Self-Attention mechanism is added to the model.Compared with multi-task learning,the experimental results have improved significantly.In terms of evaluation,the common accuracy rate,recall rate and F1 value of the named entity are used as evaluation indicators.Experiments show that the three migration learning methods are better than the traditional named entity recognition method.