Knowledge Embedding Representation Based On Vector Space Mapping

With the growing presence and demand for large-scale knowledge graphs,the need for effective knowledge embedding representations has become increasingly pressing.This thesis proposes a viable solution to knowledge embedding by introducing several key components:a pre-training based high-dimensional word vector space mapping model,a vector space mapping algorithm for structured knowledge embedding representations,and techniques for link prediction along with optimization methods.The goal is to enhance the performance of knowledge embedding representation models.In Chapter 3,we introduce a pre-trained high-dimensional word vector space mapping model that utilizes the BERT model to encode knowledge from the knowledge graph into vector representations enriched with contextual features.This approach enables the model to handle unknown and compound words more effectively through subword lemmatization,while also capturing semantic information between entities and relations using embedding extraction and pooling techniques.To enhance the accuracy and interpretability of knowledge representation,we employ a cross-entropy loss function to learn contextual information between entities and relations.Chapter 4 focuses on designing a vector space mapping algorithm specifically tailored for structured knowledge embedding representations.By utilizing a variant of the cosine similarity model,this algorithm achieves semantic matching by mapping entities and relations to a complex word vector space.Spatial locations are employed to describe the semantic features,effectively addressing the issue of data scarcity.Mathematical vector spaces are utilized to adequately represent structured knowledge information,and scoring functions are introduced.In Chapter 5,we delve into optimizing link prediction based on the knowledge embedding representation algorithm.This optimization is accomplished through improving the sampling method,rewriting the loss function,and employing adaptive parameter optimization.The sampling methods incorporate ranking-based negative case sampling and score distribution-based negative case sampling.The loss functions encompass linear combinatorial cross-entropy loss functions and loss functions derived from structured knowledge embedding models.These optimization techniques enhance the accuracy and effectiveness of link prediction,ultimately leading to improved knowledge integrity.The proposed methods in this paper provide an effective solution for knowledge embedding representation and knowledge completion tasks.They significantly improve the accuracy and efficiency of these tasks while offering enhanced support for related applications.Chapter 3 establishes a solid foundation by pre-training the spatial mapping model to learn upper and lower representations of entities and relationships and generating embedding vectors.Chapter 4 further enhances the framework by exploring structured semantic representation,mapping entities and relations into a complex word vector space to augment the model’s expressive power and semantic information capture.Finally,Chapter 5 introduces link prediction,which refines the knowledge embedding framework by optimizing sampling methods,loss functions,and parameter optimization.The interconnection and coherence among these chapters culminate in a comprehensive and effective solution.The resulting framework establishes a complete knowledge embedding framework that caters to the field of knowledge representation and link prediction,providing a holistic solution.