Multimedia Knowledge Graph Representation Learning Based On Multi-perspective Perception

Knowledge graphs(KGs)are popular data structures for representing factual knowledge to be queried and used in downstream applications.KG is typically a multirelational graph composed of(head entity,relation,tail entity)triples.Although such triples are effective in organizing structured facts,their underlying symbolic nature makes them difficult to manipulate by most machine learning algorithms.To this end,knowledge graph representation learning aims to embed symbolic entities and relations into low-dimensional continuous vector spaces,which can capture inherent structures of entities and relations and has thus quickly gained intensive attention.KG embedding models have provided an efficient and systematic solution to various real-world knowledge-driven tasks,such as relation extraction,question answering,information retrieval,and recommendation system.However,most of the current knowledge graph representation learning focuses on dealing with the entities and relationships of triples independently,so it is unable to capture the global hidden information around the neighborhood of triples;In addition,most of the methods also ignore the multirelationship neighborhood heterogeneity of the concerned entity and the information of high-order connection structure,which leads to the failure to capture more accurate semantic representation of the entity.In this thesis,we propose two new learning methods of knowledge graph representation to solve the above two problems.The main research works are as follows:(1)We propose a new end-to-end knowledge graph completion method called GAEAT(Graph Auto-encoder Attention Network Embedding)which can encapsulate both entity and relation features to solve the problem that the previous knowledge graph embedding methods do not consider the global information of the neighborhood.Specifically,we construct a triple-level auto-encoder by extending graph attention mechanisms to obtain latent representations of entities and relations simultaneously.To justify our proposed model,we evaluate GAEAT on two real-world datasets.The experimental results demonstrate that GAEAT can outperform state-of-the-art KGE models in knowledge graph completion task,which validates the effectiveness of GAEAT.(2)Considering the heterogeneity and high-order connectivity of the neighborhood of the knowledge graph,we propose a new approach for knowledge graph embedding named Contrastive Multi-relational Graph Neural Network(CMRG),which can encapsulate comprehensive features from local multi-relational triples and high-order structures of an entity.Specifically,CMRG contrasts encodings from multirelational local neighbors and high-order connectivities to obtain latent representations of entities and relations simultaneously.Experimental results demonstrate that CMRG can effectively model the multi-scale structures in KGs,and significantly outperforms existing state-of-the-art methods on benchmark datasets for the tasks of link prediction and triple classification.