POS-Tag-Based Sparse Attention Masks Reduce Transformer Compute with Linguistic Structure | HACKOBAR_