SelfElicit: Your Language Model Secretly Knows Where is the Relevant Evidence