Systems represent text or speech numerically and learn or encode patterns needed for tasks such as classification, translation, extraction, and dialogue. Modern approaches often use pretrained language models combined with task-specific data and controls.