A technique used to analyze the mistakes made by a language model, with the aim of improving its performance.