A training approach where all NLP tasks are framed as converting input text to output text, allowing a single model to handle translation, summarization, classification, and other tasks.