From Hypothesis to Premises: LLM-based Backward Logical Reasoning with Selective Symbolic Translation