Data-Oriented Parsing (DOP) models embody the assumption that humans produce and interpret natural language utterances by invoking representations of their concrete past language experience, rather than the rules of a consistent and non-redundance competence grammar. DOP models therefore maintain large corpora of sentences with syntactic structures. They analyze new input-sentences by combining partial structures from the corpus, and employ the occurrence frequencies of these structures to estimate which of the resulting analysis are the most probable one. During this seminar we will have a closer look to the computational and linguistic aspects of DOP. Recently, first Data-Oriented methods for natural language generation have been proposed, which we will discuss at the end of the seminar.