KG-BERT applies the BERT transformer architecture to knowledge graph completion tasks by treating entity and relation text as token sequences. The model performs triple classification to verify if a head-relation-tail tuple is valid, relation prediction to identify missing relationships between entities, and link prediction to infer missing links in knowledge graphs. It trains on benchmark datasets including WN11, FB13, FB15K, WN18RR, and UMLS using text representations of entities and relations.