A Machine Learning Approach to Relational Noun Mining in German

In this paper I argue in favour of a collocation extraction approach to the acquisition of relational nouns in German. We annotated frequency-based best lists of nounpreposition bigrams and subsequently trained different classifiers using (combinations of) association metrics, achieving a maximum Fmeasure of 69.7 on a support vector machine (Platt, 1998). Trading precision for recall, we could achieve over 90% recall for relational noun extraction, while still halving the annotation effort. 1 Mining relational nouns: almost a MWE extraction problem A substantial minority of German nouns are characterised by having an internal argument structure that can be expressed as syntactic complements. A non-negligeable number of relational nouns are deverbal, inheriting the semantic argument structure of the verbs they derive from. In contrast to verbs, however, complements of nouns are almost exclusively optional. The identification of relational nouns is of great importance for a variety of content-oriented applications: first, precise HPSG parsing for German cannot really be achieved, if a high number of noun complements is systematically analysed as modifiers. Second, recent extension of Semantic Role Labeling to the argument structure of nouns (Meyers et al., 2004) increases the interest in lexicographic methods for the extraction of noun subcategorisation information. Third, relational nouns are also a valuable resource for machine translation, separating the more semantic task of translating modifying prepositions from the more syntactic task of translating subcategorised for prepositions. Despite its relevance for accurate deep parsing, the German HPSG grammar developed at DFKI (Müller and Kasper, 2000; Crysmann, 2003; Crysmann, 2005) currently only includes 107 entries for proposition taking nouns, and lacks entries for PP-taking nouns entirely. In terms of subcategorisation properties, relational nouns in German can be divided up into 3 classes: • nouns taking genitival complements (e.g., Beginn der Vorlesung ‘beginning of the lecture’, Zerstörung der Stadt ‘destruction of the city’ ) • nouns taking propositional complements, either a complementiser-introduced finite clause (der Glaube, daß die Erde flach ist ‘the belief that earth is flat’), or an infinitival clause (die Hoffnung, im Lotto zu gewinnen ‘the hope to win the lottery’), or both • nouns taking PP complements In this paper, I will be concerned with nouns taking prepositional complements, although the method described here can also be easily applied to the case of complementiser-introduced propositional complements.1 In fact, I expect the task of mining relational nouns taking finite propositional complements to be far easier, owing to a reduced ambiguity of the still relatively local complementiser