zur Erlangung des akademischen Grades

This thesis discusses ways to employ a certain mathematical characterization of similarity, kernel functions, and machine learning techniques building on it to abstract from data-oriented models of language. A prominent task in machine learning and in linguistic modeling is the classification of data items, the recognition of grammatical features. A bit more complex is the modeling of surface-to-surface relations, e.g. in inflectional paradigms. These morphological changes reflect the underlying grammatical features. Traditional rule-based approaches to feature recognition and inflection fail to explain local reg-ularities within the so-called exceptions. Many linguistic domains exhibit islands of reliability (Al-bright 2002) both within classes of exceptions and the regular cases. For its best-studied instance, the debate on the proper treatment of this phenomenon is known as the past tense debate (see Pinker and Ullman 2002). More recent models (along with the increase in available computational power) adapted data-orientation, first in the shape of prototypes, and later as full-fledged exemplar models. These mode