We present a novel noisy channel model for correcting text produced by English as a second language (ESL) authors. We model the English word choices made by ESL authors as a random walk across an undirected bipartite dictionary graph composed of edges between English words and associated words in an author’s native language. We present two such models, using cascades of weighted finitestate transducers (wFSTs) to model language model priors, random walk-induced noise, and observed sentences, and expectation maximization (EM) to learn model parameters after Park and Levy (2011). We show that such models can make intelligent word substitutions to improve grammaticality in an unsupervised setting. 1