We evaluated a deep processing approach and a POS n-gram-based approach to the automatic detection of common grammatical errors in a BNC-derived artificial error corpus. The results are broken down by error type. Together with the deep approach, a decision tree machine learning algorithm can be used effectively. However, extending the shallow approach with the same learning algorithm gives only small improvements. Combining the deep and shallow approaches gives an additional improvement on all but one error type.
Our plan is to investigate why all methods perform poorly on missing word errors, to extend the error creation procedure so that it includes a wider range of errors, to try the deep approach with other parsers, to integrate additional features from state-of-the-art shallow techniques and to repeat the experiments for languages other than English.