Research On Named Entity Recognation Based On Transfer Learning

Named Entity Recognition task is a basic task of Natural Language Processing.With the development of information,mining information has become very critical.Improvements in hardware technology are making deep learning shine in Artificial Intelligence.Named Entity Recognition aims to identify entity words from unstructured text.In recent years,many advanced models use transfer learning technology to complete the knowledge transfer between different tasks and different fields,thus enhancing the ability of model feature extraction.However,there are still some problems in the entity recognition model based on transfer learning,such as noise generated in the process of feature fusion,lack of labeled data,and imbalance of samples.Task shared features and domain shared features of Sequence Labeling are extracted efficiently by improving transfer learning technology,to improve the entity recognition effect.The paper mainly includes the following two research points:Firstly,this paper proposes Adversarial Named Entity Recognition with POS label embedding to solve the problem of noise generated in the process of feature fusion.To capture part-of-speech dependencies,the paper first employs POS label embedding to represent part-of-speech features and the Self-Attention mechanism is adopted to capture dependencies between different words at any position in the sentence.To distill tasks-shared features and suppress task-specific noise,the paper takes advantage of the adversarial training and task-attention mechanism to map tasks-shared information between POS and NER to the shared feature space.For noun constraints,the paper adds the constraint-based on the CRF formula of the NER task to improve the recall of the model.The experimental results show that the proposed method can effectively extract the part-of-speech features and assist the named entity recognition task.Moreover,this paper proposes Multi-domain Named Entity Recognition Based on adversarial meta-learning to solve the problem of the lack of labeled data,and imbalance of samples.Model by combining the meta encoder and the discriminator,the meta-model learning to share the best initial point and domain shared features,thus effectively complete the knowledge transfer.As well as the sub-model can converge through a few times iteration in the target domain,reduce the label data dependence in the target domain.In addition,for the problem of unbalanced samples,this paper uses different loss functions for in-depth exploration.The experimental results show that the model effectively learns the common initial point and the shared features of the domain,so as to solve the Named Entity Recognition in the low-resource domain.