Using language models to generate structured information from text rather than identifying fixed spans, allowing more flexible output formats.