Weakly Supervised Grammatical Error Correction using Iterative Decoding
This addresses grammatical error correction for language learners, with incremental improvements in weakly supervised methods.
The paper tackles grammatical error correction by using weakly supervised bitext from Wikipedia revisions and iterative decoding, achieving an F0.5 of 58.3 on CoNLL'14 and GLEU of 62.4 on JFLEG, with an F0.5 of 48.2 without labeled data.
We describe an approach to Grammatical Error Correction (GEC) that is effective at making use of models trained on large amounts of weakly supervised bitext. We train the Transformer sequence-to-sequence model on 4B tokens of Wikipedia revisions and employ an iterative decoding strategy that is tailored to the loosely-supervised nature of the Wikipedia training corpus. Finetuning on the Lang-8 corpus and ensembling yields an F0.5 of 58.3 on the CoNLL'14 benchmark and a GLEU of 62.4 on JFLEG. The combination of weakly supervised training and iterative decoding obtains an F0.5 of 48.2 on CoNLL'14 even without using any labeled GEC data.