Musical schemata constitute important structural building blocks used across historical styles and periods. They consist of two or more melodic lines that are combined to form specific successions of intervals. This paper tackles the problem of recognizing voice-leading schemata in polyphonic music. Since schema types and subtypes can be realized in a wide variety of ways on the musical surface, finding schemata in an automated fashion is a challenging task. To perform schema inference we employ a skipgram model that computes schema candidates, which are then classified using a binary classifier on musical features related to pitch and rhythm. This model is evaluated on a novel dataset of schema annotations in Mozart's piano sonatas produced by expert annotators, which is published alongside this paper. The features are chosen to encode music-theoretically predicted properties of schema instances. We assess the relevance of each feature for the classification task, thus contributing to the theoretical understanding of complex musical objects.