Defending against Backdoor Attacks in Natural Language Generation