FAQに戻る
Enterprise Applications

What is perplexity

Perplexity measures how uncertain a language model is when predicting the next word in a sequence. It quantifies how well a model predicts a sample of text or data; lower values indicate better predictive performance.

A key principle is that perplexity is directly tied to the model's calculated probability of the test data. It is calculated as the exponential of the cross-entropy loss. Crucially, lower perplexity signifies the model is less "perplexed" by the data and makes more confident predictions. Comparisons are only valid between models evaluated on the exact same dataset and vocabulary. It primarily assesses generative capabilities for sequential data like text.

Perplexity is vital for evaluating, comparing, and selecting language models during development and research. It provides an objective, quantifiable metric for model performance before costly human evaluation. It helps researchers identify improvements and optimize models for specific tasks like machine translation or text generation.

関連する質問