Machine learning and statistical methods, such as unsupervised semantic models, make massive cultural heritage collections more explorable and analyzable. These models capture many underlying patterns of raw textual and visual materials, but neither model creators nor model users fully understand which specific patterns are learned by a given model nor under what conditions a particular pattern becomes more learnable. In this dissertation I address two core questions (i) what do models actually learn? and (ii) how can we direct what they learn? Instead of proposing new models, I focus on expanding the affordances, as well as our understanding, of existing ones that are used by scholars in the humanities and social sciences. In the first part of this dissertation, I study what models learn by way of expanding the ways in which they can be used. In the second part, I investigate how existing models can be directed away from known, uninteresting structures via corpus- and representation-level interventions. Throughout this work, I show how machine learning and statistical methods provide an opportunity to view collections from alien, defamiliarized perspectives that can call into question the boundaries of established categories. Likewise, I show how the uses of computational methods within humanities and social science scholarship can test, challenge, and expand the affordances of these methods. Ultimately, this dissertation highlights some of the many ways in which machine learning and the humanities help one another.