A training technique where random words in text are hidden and the model learns to predict them, commonly used in models like BERT.