Unlearning Reveals the Influential Training Data of Language Models