A　knowledge　graph　is　a　knowledge　base　of　graph　structure　that　stores　massive　facts　and　their　relations　in　the　form　of　triplets.　Embedding　entities　and　relations　in　knowledge　graph　into　low-dimensional　continuous　vector　space　is　of　great　significance　for　downstream　tasks,　whereas　most　existing　methods　give　an　end-to-end　fixed　representation.　Based　on　deep　metric　learning,　we　give　different　embeddings　for　the　same　entity　in　different　relations　and　positions,　thus　dynamically　representing　entities　in　knowledge　graph　and　effectively　modeling　complex　relations.　In　addition,　a　path-based　negative　sampling　method　is　proposed　to　widen　the　distance　between　similar　entities　to　better　deal　with　the　aggregation　problem　of　embedding　entities　in　the　representation　space　under　one-to-many　relation.　Experiments　show　that　the　MRR　and　Hit@1　of　this　method　on　the　linked　prediction　dataset　FB15k-237　reach　0.360　and　0.267　respectively,　which　are　higher　than　the　mainstream　distance　models　and　semantic　matching　models.　Finally,　the　MRR-Dimension　curves　demonstrates　that　our　model　can　perform　well　in　both　low　and　high　embedding　dimensions　and　has　the　potential　to　be　applied　to　large-scale　knowledge　graph.　©　2023　Copyright　held　by　the　owner/author(s).