“Surprise” is used as a metaphor to help explain Perplexity. A model with higher perplexity is said to be more surprised by the output. This is because the model is more likely to choose the next word that is less correct.