Enhancing elusive clues in knowledge learning by contrasting attention of language models