Vocabulary knowledge is essential to educational progress. High quality vocabulary instruction requires supportive contextual examples to teach word meaning and proper usage. Identifying such contexts by hand for a large number of words can be difficult. In this work, we take a statistical learning approach to engineer a system that predicts informativeness of a context for target words that span the range of difficulty from middle school to college level. Our database (released open source) includes 1,000 hand-selected words associated with approximately 70,000 contextual examples gathered from the Internet. Our training data included each context rated by 10 individuals on a four-point informativeness scale. We process the text of each context into a novel collection of approximately 600 numerical features that captures diverse linguistic information. We then fit a nonparametric regression model using Random Forests and compute out-of-sample prediction performance using cross-validation. Our system performs well enough that it can replace a human judge: for a target word not found in our dataset, we can provide curated contexts to a student learner such that most of the contexts (54 percent) feature rich contextual clues and confusing contexts are rare (<1 percent). The quality of our curated contexts was validated by an independent panel of high school language arts teachers.