Convolution kernels support the modeling of complex syntactic information in machine-learning tasks. However, such models are highly sensitive to the type and size of syntactic structure used. It is therefore an important challenge to automatically identify high impact sub-structures relevant to a given task. In this paper we present a systematic study investigating (combinations of) sequence and convolution kernels using different types of substructures in document-level sentiment classification. We show that minimal sub-structures extracted from constituency and dependency trees guided by a polarity lexicon show 1.45 point absolute improvement in accuracy over a bag-of-words classifier on a widely used sentiment corpus.