We address the tasks of recovering bibliographic and document structure metadata from scholarly documents. We leverage higher order semi-Markov conditional random fields
to model long-distance label sequences, improving upon the performance of the linear-chain conditional random field model. We introduce the notion of extensible features, which allows the expensive inference process to be simplified through memoization, resulting in lower computational complexity. Our method significantly betters the state-of-the-art on three related scholarly document extraction tasks.