Chinese Entity Recognition And Relation Extraction Based On Deep Learning

With the development of the Internet,every user of the Internet has become the creator of content.In order to obtain the content from the massive information quickly,the information extraction technology is used.The main purpose of information extraction is to transform unstructured text into structured text.There are two basic tasks in Chinese information extraction: Chinese named entity recognition(NER)and named entity relation extraction(RE).The purpose of Chinese NER is to identify and mark named entities such as person names,locations and organizations in Chinese text.The purpose of entity relation extraction is to get the relationship between entity pairs.By using entity pairs and their relationship,entity relation triples can be formed to represent knowledge.There are two methods of Chinese named entity and relationship joint extraction: pipeline model and joint model.This paper realized Chinese named entity and relationship extraction by these two methods,and improved the result of the models.In the task of Chinese NER,this paper adopts the Bilattice-LSTM-ATT-CRF model.The model uses the Bilattice-LSTM layer to semantically encode the input text.Then the model uses the attention mechanism to weight the semantically encoded vector,and finally decodes the vector through CRF to realize Chinese NER.In this model,the precision is 95.64%,the recall is 93.67%,and the F value is 94.44%.DP-BiLSTM-SelfATT model was used in the task of relationship extraction.The model used dependency parsing as extra semantic information.In this model,the precision is 78.96%,the recall is 47.43%,and the F value is 59.26%.In the joint model,this paper transforms the joint extraction task into sequence annotation task with a special labeling strategy.Based on the BiLSTM-LSTM end to end model,POS and dependency parsing are used as extra information.Attention mechanism and objective function with bias term are used in this model to realize the joint extraction.In this model,the precision is 76.23%,the recall is 44.53%,and the F value is 56.22%.