This paper introduces a new extensible architecture for defining the metadata of multimedia from different sources, extracting this metadata from documents, and representing it to users. We introduce meta-metadata, semantic data structures that guide the extraction and manipulation of strongly typed metadata, including visual representations, from diverse documents. Meta-metadata declarations are automatically translated into metadata class definitions. Both are defined using the ecologylab.xml information binding framework. Extensions to the framework support manipulation of instances of the generated metadata classes with generic field accessor objects, enabling information extraction, information visualization, contextual metadata presentation, editing, and interaction. We show how meta-metadata and the metadata it generates are used in the mixed-initiative information composition information discovery support tool, combinFormation.