position representation is crucial for building position-aware representations in transformers. Existing position representations suffer from a lack of generalization to test data with unseen lengths or high comp