Tuning a Grammar Correction System for Increased Precision

In this paper, we propose two enhancements to a statistical machine translation based approach to grammar correction for correcting all error categories. First, we propose tuning the SMT systems to optimize a metric more suited to the grammar correction task (F-β score) rather than the traditional BLEU metric used for tuning language translation tasks. Since the F-β score favours higher precision, tuning to this score can potentially improve precision. While the results do not indicate improvement due to tuning with the new metric, we believe this could be due to the small number of grammatical errors in the tuning corpus and further investigation is required to answer the question conclusively. We also explore the combination of custom-engineered grammar correction techniques, which are targeted to specific error categories, with the SMT based method. Our simple ensemble methods yield improvements in recall but decrease the precision. Tuning the custom-built techniques can help in increasing the overall accuracy also.