Text-enhanced Knowledge Graph Representation Learning Based On Graph Attention Network

Knowledge Graph Representation Learning(KGRL),also known as Knowledge Graph Embedding(KGE),aims to mine the valid information contained in the knowledge graph and transform it into vector representation,so that it can be more easily applied to machine learning and deep learning models to achieve knowledge retrieval,knowledge inference and other knowledge application purposes.Improving the expressive ability of the Knowledge Graph Representation Learning model mainly includes two aspects.1)Improve the model architecture.Based on the traditional KGRL model,combining the advantages of Convolutional Neural Network models,Attention Mechanisms,and Graph Neural Network models,a new model architecture is optimized.2)Expand the knowledge base.Due to the graph information cannot be completely modeled only through the structural data of the knowledge graph,an attempt is made to expand the knowledge graph through some external knowledge base information,and design a corresponding knowledge enhancement model to fully mine the content information of the knowledge graph.The main research contents of this article includes: 1)We optimized the KGRL algorithm in view of the above two aspects and proposed a Text-Enhanced Knowledge Graph Representation Learning Based on Graph ATtention Network,TGAT for short.On this basis,we verified the TGAT model's performance on the link prediction task of WN18 RR and FB15K-237 datasets.2)In order to verify the performance of the algorithm in real scenarios,we combined the Insurance's underwriting mission scenarios,constructed a medical knowledge graph dataset MI225,and verified the performance of the TGAT model on the MI225 dataset.3)In order to study the organic combination of KGRL models and deep learning models,we proposed a new model based on KGRL and the Attention Mechanism for specific underwriting tasks,named KDBAN.Experiments showed that the TGAT model improves significantly on metrics such as MR,MRR,and Hits@N.At the same time,the KBDAN model with the addition of TGAT performs better than other classic KGRL models.In addition,the newly proposed deep underwriting model KBDAN based on knowledge representation is a significant improvement over various classical machine learning frameworks and deep learning models for classification tasks.This paper systematically implements the knowledge extraction of structural information and content information of the knowledge graph.Classification and recommendation skills were improved by applying the pre-trained knowledge representation model to the deep learning task.