Semantic Consistency for Assuring Reliability of Large Language Models