Evaluating the Factual Consistency of Large Language Models Through News Summarization