This paper describes our submission to the CoNLL 2014 shared task on grammatical error correction using a hybrid approach, which includes both a rule-based and an SMT system augmented by a large webbased language model. Furthermore, we demonstrate that correction type estimation can be used to remove unnecessary corrections, improving precision without harming recall. Our best hybrid system achieves state of-the-art results, ranking first on the original test set and second on the test set with alternative annotations.