A Study On Entity Relation Extraction Based On Deep Learning

Google put forward a pre-training model for BERT(Bidirectional Identification from Transformers)in 2018,And in a number of NLP tasks to achieve outstanding results.The BERT-CRE(BERT-Chinese-Relation-Exraction)model designed by combining BERT with Chinese full-supervision relation extraction,According to the position of the word in different sentences,Using the relation,Dynamic generation of 768 dimensional vector representation,A more semantic expression,And can improve the extraction effect and speed.The experimental results show that the F1 value of the BERT-CRE is increased from 65.61% to 73.35%compared with the baseline model,AUC increased from 57.33% to 64.72%,The convergence rate is increased by nearly 6 times.To reduce the dependency of fully supervised data sets on manual data tagging,distant supervision aligns unlabeled databases with existing knowledge bases,Placing an entity on all the same instances in a triple package of entities and relations,Assuming that the relation of the entity pair in the knowledge base is the relation of the package,Transforming traditional manual marking into automated alignment through strong assumptions,Save a lot of labor costs.BERT-ATT(BERT-Attention)model combining remote supervision with BERT-based English relations extraction.Distant supervision of data,The experimental results show that the F1 value of BERT-ATT compared with the baseline model increased from 77.4% to 87.23%,AUC increased from 84.48% to 93.73%,Precision increased from 64% to 66%,The convergence rate is nearly 2 times higher.Differences between languages lead to different language descriptions of the same relation,but the semantics are the same.In order to reduce the noise effect of distant supervision,the Chinese and English multilingual model MNRE-ADV(Neural Multi-lingual Neutral Relation Extraction with Adversarial Training)is extracted from the distant supervisied relation.Align the same entity in different languages,construct multilingual data set and train model to find consistent semantic space.Compared with the single language model,the AUC value increased from 32.3%(Chinese)and 35.7%(English)to 47.4%(Chinese and English),which achieved cross-language feature fusion and reduced noise effect.By introducing noise and discriminator,the AUC value is increased from 41.3% to 46.2%,which enhances the robustness and generalization ability of the model.In order to verify the effectiveness of the method,a set of entity relation extraction tools and relational map display query tools are constructed based on the models.The entity relation extraction tool visualizes the relation extraction model,and the user obtains the predicted relation category by providing the entity pair and the sentence.The relation graph display query tool realizes the display and query of the knowledge base of the existing relation and entity,displays the complex relation in the form of map,and provides inspiration for the analysis of information research and analysis.