Using Prior Knowledge to Guide BERT’s Attention in Semantic Textual Matching Tasks