Multilingual Knowledge Graph Completion with Negative Sample Balance Based Adaptive Self-supervised Graph Alignment

Due to the high cost of manual annotation, knowledge graph reasoning and completion have always been important research directions in this field. Current research often focuses on monolingual knowledge graphs. Multilingual knowledge graphs often share real-world entities and relationships, serving as bridges to harmonize different knowledge graphs. This not only helps disseminate knowledge to low-resource knowledge graphs but also alleviates the need for expensive manual annotation. Therefore, research on multilingual knowledge graph reasoning is crucial. We study knowledge reasoning in multilingual knowledge graphs, proposing a model that combines negative sample balancing with self-supervised graph alignment. Main contributions include: (1) Using K-means clustering for entity embeddings to categorize entities and improve negative sample sampling, with a weighted negative sample generation method for balancing. (2) Leveraging relation-aware and linguistic information for knowledge representation learning across languages, followed by entity alignment based on this representation. Iteratively identifying new alignments in a self-supervised manner. Finally, experiments on typical multilingual knowledge graphs, EPKG and DBP5L, validate the effectiveness of our method.