Research On Open Relation Extraction Based On Deep Learning

Relation Extraction(RE)is a basic task in Natural Language Processing(NLP),which supports downstream tasks such as information retrieval and knowledge graph construction.However,traditional Relation Extraction is usually closed,which can not extract the new relation in the text.Therefore,as an improvement method,Open Relation Extraction(Open RE)is proposed to extract relational facts from open domain corpus without pre-defining relational types,and its task setting has irreplaceable advantages compared with qualified Relation Extraction.In order to improve the accuracy of Open RE and realize the automatic extraction of factual knowledge from text,this paper adopts the deep learning method,and mainly does the following work:Firstly,the existing Open RE methods based on clustering usually adopt unsupervised mode,ignoring the advantages of supervised learning,this paper proposed an unsupervised ensemble clustering framework.In this method,unsupervised ensemble learning is combined with multi-step clustering algorithm based on information metric to independently create high-quality pseudo labels,which is used as supervised information to improve the learning of relational features,thus guiding the clustering process to obtain better label quality.Finally,through multiple iterative clustering,the relationship types in the text can be effectively discovered.Experimental results on Few Rel and NT-FB datasets show that the proposed method is superior to other mainstream Open RE models,with F1 values reaching 65.2% and 67.1%,respectively.Then,in view of the existing methods only consider the information of the instance itself,but ignore the knowledge of the relation between the instances,this paper studies the graph convolution autoencoder method based on cross-attention fusion.Specifically,a new relational representation embedding scheme is proposed,which uses the type characteristics of entities to enhance the richness of the representation.Secondly,the mechanism of parallel learning of sentence instance and semantic similarity between sentences is proposed,and the attention mechanism is used for cross fusion and propagation update.Finally,a more efficient and automatic clustering algorithm,Louvain algorithm,is used to form the relation types.The results of several experiments show that the proposed method is effective and advanced.Finally,from the practical application,the prototype system of open relation extraction is developed,the proposed open relation extraction model is packaged,the B/S architecture model is used for system design,and Easy UI+Spring Boot+Mysql technology is used for system development,so as to visualize the results of open relation extraction task.