Making the Most of Text Semantics to Improve Biomedical Vision–Language Processing