Triggerless Backdoor Attack for NLP Tasks with Clean Labels