Massive knowledge graphs (KGs) are increasingly significant in modern information systems. Previous research on knowledge graph completion must gather sufficient training instances for new introduced relations in order to increase the coverage of KGs. This study examines an innovative approach to the scarce labelling problem. The model use text descriptions to learn semantic elements of newly added relations, allowing us to recognise facts even when there are few examples available. To learn relations based on text, we utilised a GAN-BERT model, which is a modified form of BERT. Four distinct datasets are employed for this. Experiments demonstrate that using GAN-BERT decreases the demand for annotated instances and improves performance in various relation prediction tasks.