Are LLMs Really Not Knowledgable? Mining the Submerged Knowledge in LLMs' Memory