Make LLMs better zero-shot reasoners: Structure-orientated autonomous reasoning