Evaluating and Improving a Derivational Lexicon with Graph-theoretical Methods

We employ a graph-theoretical approach to evaluate and improve a German derivational lexicon, DERIVBASE. We represent derivational families (that is, groups of derivationally related words) as labelled directed graphs in which words (friend, friendly) are nodes and derivational relationships (friend → friendly) between words are directed edges, labeled with the derivation rule (-ly). This graph-theoretical approach allows us to carry out a large-scale comparison of the structure of different derivational families and identify, in a completely automatic fashion, possible errors in the resource. We conduct a manual evaluation of the predictions of our method and find that it successfully identifies instances which are missing from DERIVBASE; the predictions of our approach can be interpreted as the result of interplay among productivity constraints.