Evaluating Models of Latent Document Semantics in the Presence of OCR Errors