Extracting Rule-based Descriptions of Attention Features in Transformers