Text-based Entity And Relation Extraction And Representation And Reasoning Of Knowledge Graphs

With the rapid development of information and communication technologies,especially the mobile internet.Human beings are gradually entering the era of data explosion,and there are a large amount of data and knowledge generated every day.Faced with massive data,how quickly and effectively users get the used knowledge from data has become a hot topic in this era.The storage and representation of massive unstructured text data in the form of knowledge graphs has become a mainstream way of knowledge storage.In this process,extraction of entities and relations from the text and the reasoning and representation of existing knowledge graphs become the key technologies for better application of knowledge graphs in specific fields.Therefore,the research content of this paper is as follows:1.The existing joint extraction methods of entities and relations are still too simple in the encoding module to fully represent the context semantics.For the limitation of the encoding module,the proposed joint extraction method in this paper add the multi-head attention and concatenate the multi-module output semantics vector in the encoding module to enrich the semantics representation of the text.Based on the rich semantics features,the method can get better performances.2.In the numerous knowledge representation methods,the translation methods as one of the mainstream methods perform well in many public datasets.Similar to other embedding methods,the translation methods also embed entities and relations into the low-dimensional continuous tensor space.However,the existing translation methods can not represent the complex relations accurately,such as one-to-many,many-to-one,many-to-many relations.For the limitations of the baseline representation methods,a novel asymmetric embedding method for knowledge graphs completion(AEM)is proposed.In the AEM method,the head entities and tail entities are projected to two relational sub-vector space in the same relation space.The representation problem of complex relations is relieved to a great extent,at the same time,the representation method becomes easy to train and easy to use.3.To solve the issue of the AEM method that cannot flexibly represent the triple relations,an asymmetric knowledge representation learning in manifold space(MAKR)is proposed inspired by the OrbitE.The proposed MAKR embeds the tail entities projected to the relational space into the hyper-dimensional manifold sphere instead of a point in the corresponding relational space.Therefore,the proposed MAKR can alleviate the loose entity representation issue in the relational space greatly.In this paper,the proposed MAKR is trained and tested on the open dataset.Compared to the baselines,the experimental results are significantly improved which proves the validity of the proposed MAKR in this paper.The above research contents cover two sub-tasks in the knowledge graph,joint entity and relation extraction and knowledge graphs representation and reasoning.Based on the shortcomings of the baseline methods in the two sub-tasks,the improved method proposed in this paper has achieved better improvement effect,which can certain reference significance for the technical progress in the field of knowledge graphs.