This study presents an approach to the automatic derivation of structured linguistic descriptions from a set of isolated individual data. Structure inferencing takes place on the basis of the identification of relations between the data, plus generalizations about those data. Quality criteria are used for the formal evaluation of various descriptions for a given set of data and the direct comparison of those descriptions. This provides the basis for the selection of good descriptions. The efficiency of the procedure is illustrated by application to linguistic data relating to noun inflexion and syntactic verb classification in German.