| Named Entity Recognition(NER)is one of the classical tasks in Natural Language Process(NLP).NER is utilized to identity entities with specific meaning like person,location and specific domain term from text and the downstream NLP tasks is able to get precise entity’s semantic information through NER,which makes some fine-grained downstream NLP tasks reach great performance.In the present,there is a research paradigm in Chinese NER research.Such paradigm models text as sequence data and extracts semantic information via sequence models like RNN and LSTM and finally,it identifies entities from semantic information.To make Chinese NER performance better,it also develops the data enhancement method with other textual features such as character,grammar etc.More importantly,with the development and implement of pre-train models in NER,the common semantic information,learned by pre-train models from mass data,enriches the understanding capacity of Chinese NER models.These pre-train models improve the performance of Chinese NER greatly.In our opinion,it’s useful to aid Chinese NER performance with the help of pretrain tasks.But as for Chinese NER research,there’re still a lot of problems to be solved.In present,Chinese NER research paradigm contains 2 problems which limits Chinese NER performance.The first one is long-distance contextual information.As for Chinese,to understand word meaning precisely,it’s necessary to combine word meaning and its context.But the present paradigm labels text as a form of sequence data,which it focuses more on the semantic information in word itself and its’ adjoin words and neglects the long-distance contextual information.The second problem is the iteration between semantic features and other features.The data enhancement in Chinese NER is the iteration between semantic features and other features like grammar,textual etc.It’s still rough for present data enhancement to get better performance.There is also a great gap between Chinese and English that it is not suitable to transfer the data enhancement methods from English NER research to Chinese NER research directly.With the great performance of GNN(Graph Neural Network)recently,a lot of researchers focus more on GNN.GNN is able to capture topology features through data after constructing graph,which can get better performance in semi-structured data form’s model.To tackle two above problems,this paper labels text as a form of semistructured data and proposes a GNN-based Chinese NER model.This paper also proposes a features iteration adapted to this GNN-based Chinese NER model.Specifically,this paper’s work constructs with two parts:1.For the long-distance contextual information,this paper proposes a LEC-Net(Lexicon Enhanced Collaborative Graph network).LEC-Net not only extracts semantic features that it focuses on textual semantic information through its sequence model,but also extracts graph topology features that it focuses on long-distance context information by labeling text as semi-structure data.2.For the feature iteration,this paper designs a feature iteration method between the extracted semantic features and graph topology features from LEC-Net.This iteration method can make full use of semantic information and contextual information than plays a role as data enhancement.This iteration method better model performance in Chinese NER.There is an experiment conducted in 3 Chinese NER dataset: Weibo,MSRA and CLUE to verify the effectiveness and efficiency of model proposed in paper in Chinese NER research.Compared to other models,The LEC-Net can reach about 2% greater F1 scores in the experiment.What’s more,our experiment avoids to use pre-train models to verify the effectiveness of GNN and features iteration from LEC-Net directly that it makes our research focuses more on the core of Chinese NER research.LEC-Net can be set as an example for other NLP research. |
