A Hybrid Model For Grammatical Error Correction

This paper presents a hybrid model for the CoNLL-2013 shared task which focuses on the problem of grammatical error correction. This year’s task includes determiner, preposition, noun number, verb form, and subject-verb agreement errors which is more comprehensive than previous error correction tasks. We correct these five types of errors in different modules where either machine learning based or rule-based methods are applied. Preprocessing and post-processing procedures are employed to keep idiomatic phrases from being corrected. We achieved precision of 35.65%, recall of 16.56%, F1 of 22.61% in the official evaluation and precision of 41.75%, recall of 20.29%, F1 of 27.3% in the revised version. Some further comparisons employing different strategies are made in our experi-