Musical structure plays an essential role in the listener’s perception of a piece of music. In Western tonal tradition, the structure of a musical score results from a number of factors including repetitions, variations, and typical harmonic progressions such as cadences. Cadences contribute to the segmentation of the musical discourse into successive phrases by the feeling of closure or relaxation they provide to the listener (Blombach, 1987). Modellng cadences is therefore a promising task to improve our understanding of the perception of musical structure. We present an approach to model cadences, based onautomatic corpus extraction of theory driven features that are commonly used to describe cadences. Experiments on two different corpora from J.-S. Bach and J. Haydn showt that, despite their perceptive salience and apparent univocity, cadences turn out to be challenging to model computationally and variable depending on the musical style. We will then present how the retrieval of local structural markers such as cadences can be extended to target the modeling of large-scale structures and therefore bring new modeling and cognitive challenges. This approach will be illustrated with the computational retrieval of the sonata form which has been intensively used in the classical style.