We present a method for automatic determiner selection, based on an existing language model. We train on the Penn Treebank and also use additional data from the North American News Text Corpus. Our results are a significant improvement over previous best.