An Approach to NMT Re-Ranking Using Sequence-Labeling for Grammatical Error Correction

Abstract

An approach to N-best hypotheses re-ranking using a sequence-labeling model is applied to resolve the data deficiency problem in Grammatical Error Correction (GEC). Multiple candidate sentences are generated using a Neural Machine Translation (NMT) model; thereafter, these sentences are re-ranked via a stacked Transformer following a Bidirectional Long Short-Term Memory (BiLSTM) with Conditional Random Field (CRF). Correlations within the sentences are extracted using the sequence-labeling model based on the Transformer, which is particularly suitable for long sentences. Meanwhile, the knowledge from a large amount of unlabeled data is acquired through the pre-trained structure. Thus, completely revised sentences are adopted instead of partially modified sentences. Compared with conventional NMT, experiments on the NUCLE and FCE datasets demonstrate that the model improves the F_0.5 score by 8.22% and 2.09%, respectively. As an advantage, the proposed re-ranking method has the advantage of only requires a small set of easily computed features that do not need linguistic inputs.