Unseen Relation Prediction Using BERT and Neighbour Encoder

Knowledge graphs are invaluable assets for numerous applications in computational intelligence. Recent research has focused on creating link prediction models that are capable of completing knowledge graphs. This paper suggest leveraging pre-trained language models to complete knowledge graphs for unseen relationships. Predicting unseen relations in a knowledge graph is a difficult task that requires inferring relationships between items that may not be explicitly stated in the available data.  Our technique, KG-Sim-BERT, employs relation descriptions to assess the possibility of linkage with other triplets. We use the BERT pre-trained model to represent KG triplets and assess their similarity. Experiments on FB15k-237 and Yago3-10 data sets demonstrate that our technique can predict previously unseen relations using BERT and cosine similarity.