In this paper, we describe our experiments in preposition disambiguation based on a – compared to a previous study – revised annotation scheme and new features derived from a matrix factorization approach as used in the field of distributional semantics. We report on the annotation and Maximum Entropy modelling of the word senses of two German prepositions, mit (‘with’) and auf (‘on’). 500 occurrences of each preposition were sampled from a treebank and annotated with syntacto-semantic classes by three annotators. Our coarse-grained classification scheme is geared towards the needs of information extraction, it relies on linguistic tests and it strives to separate semantically regular and transparent meanings from idiosyncratic meanings (i.e. of collocational constructions). We discuss our annotation scheme and the achieved inter-annotator agreement, we present descriptive statistical material e.g. on class distributions, we describe the impact of the various features on syntacto-semantic and semantic classification and focus on the contribution of semantic classes stemming from distributional semantics.