| Named entity classification is one of the key research fields of information extraction in recent years.It aims to classify the entities in the text.The traditional entity classification task can identify few types,while the entity types are complex and diverse in real life.More detailed classification will provide richer information for information extraction.At present,named entity classification has achieved a great breakthrough in the field of labeled data supervision.However,for specific fields,such as military,medicine and small languages,there are multiple limitations when labeling such as high cost and long time-consuming;Therefore,unsupervised entity classification has high research value.In the early stage,unsupervised entity classification mainly assisted entity classification by manually extracting features,which has poor portability.Secondly,unsupervised clustering algorithm is limited by parameters,data distribution and dimensions in entity classification task,resulting in the development of unsupervised domain lagging behind the supervised domain.Based on the above problems,this paper proposes three fine-grained classification methods based on English feature clustering,In order to obtain a variety of features to enrich semantic information and optimize the clustering algorithm for fine-grained hierarchical classification tasks,the research contents of fine-grained entity classification based on joint representation clustering proposed in this paper are as follows:(1)In order to enrich the feature information input of classification task,I propose a fine-grained entity classification method based on joint representation in this paper without any annotation data.The joint representation method aims to enrich the feature information input of the model in different various fields.It uses three kinds of representations,namely word level representation,sentence level representation and entity level representation.Based on letters,sentences and words,different feature extraction channels are used to obtain different forms of representation results,so as to enrich the feature type information of the model to improve the generalization ability of different fields of the model.Experiments show that the joint representation model can effectively improve the entity classification performance,The results show that the joint representation method in this paper has higher classification accuracy and universal applicability,and the F1 value is increased by 6% compared with the existing advanced model.(2)In view of the problem that hierarchical clustering algorithm is sensitive about data distribution and possesses high time complexity,this paper proposes a hierarchical clustering optimization algorithm based on complex distributed data.Firstly,set the merging threshold,input the three representations into hierarchical clustering,and merge them into a cluster when the similarity reaches the threshold or above.The similar features are combined through clustering,each cluster is assigned an entity level representation,and the combination is linked to the knowledge base to predict the entity category,and the longest category information path is selected through fine-grained indicators to give the whole cluster,Finally,this paper compares and analyzes this model with the latest supervision model in the biomedical data set,and realizes the competitive results with the supervision model. |
