Deep Associations, High Creativity: A Simple yet Effective Metric for Evaluating Large Language Models