Understanding Perplexity Metrics in Natural Language AI

New, state-of-the-art language models like DeepMind's Gopher, Microsoft's Megatron, and OpenAI's GPT-3 are driving a wave of innovation in NLP. How do you measure the performance of these language models to see how good they are? In a previous post, we gave an overview of different language model evaluation metrics. This post dives more deeply into one of the most popular: a metric known as perplexity. Imagine you're trying to build a chatbot that helps home cooks autocomplete their grocery shopping lists based on popular flavor combinations from social media.