Supervised Entities And Relations Joint Extraction Based On Deep Learning

With the rapid development of big data and machine learning technology,people can process and analyze more text data.However,the knowledge and information contained in text data are very rich and complex,so an automated method is needed to help us extract useful knowledge and information from it.Entity relation extraction technology is a method that can automatically extract information and knowledge,and it is also one of the key technologies for information extraction and knowledge graph construction.It can automatically identify entities in the text and infer the relationships between entities,generating semantic knowledge in the format of <subject entity,relation,object entity> triplets.Therefore,we can construct knowledge graphs through entity relation extraction technology,and expand them for better assistance in intelligent retrieval,intelligent question answering and other applications.There are two main methods for entity relation joint extraction based on deep learning: pipeline method and joint learning method.The pipeline method usually executes the entity recognition and relation extraction subtasks in sequence,but there are problems such as error accumulation,redundant entity information,and insufficient information interaction.In contrast,the joint learning method can better integrate entity and relation information,and extract entities and their corresponding relationships,effectively solving the problems in the pipeline method.Based on the joint learning method,this article focuses on the connection and difference between entity recognition and relation extraction subtasks,and proposes the following two models:1.This article proposes a joint entity relation extraction model based on2D-CNN feature fusion and propagation.In order to achieve bidirectional feature interaction propagation at the task level,two isomorphic but different direction2D-CNNs are introduced in the model,one encoding features from entity recognition to relation extraction and the other from relation extraction to entity recognition.Then,by fusing the bidirectionally encoded feature information,the entity feature representation and relation feature representation with bidirectional interaction information are obtained.The experimental results on the NYT and Web NLG datasets demonstrate that the proposed model outperforms the baseline model,with an improvement in F1 scores of 0.6 and 3.4,respectively.2.This article also proposes a joint entity relation extraction model based on an interactive biaffine mechanism.The model further refines the entity relation extraction task into three subtasks: subject entity recognition,object entity recognition,subject-object relation classification.In each subtask,the biaffine mechanism is introduced to deeply learn the differential feature information and interaction feature information between the subtasks.Extensive experiments conducted on the NYT and Web NLG datasets have demonstrated substantial improvement in the model’s F1 scores,with F1 scores reaching 93.4 and 90.6,respectively.