A unified framework for grammar error correction

Abstract

In this paper we describe the PKU system for the CoNLL-2014 grammar error correction shared task. We propose a unified framework for correcting all types of errors. We use unlabeled news texts instead of large amount of human annotated texts as training data. Based on these data, a tri-gram language model is used to correct the replacement errors while two extra classification models are trained to correct errors related to determiners and prepositions. Our system achieves 25.32% in f0.5 on the original test data and 29.10% on the revised test data.