Grammatical Error Correction through Round-Trip Machine Translation

Findings of the Association for Computational Linguistics: EACL 2023, pages 2208–2215 May 2-6, 2023 ©2023 Association for Computational Linguistics Grammatical Error Correction through Round-Trip Machine Translation Yova Kementchedjhieva University of Copenhagen yova@di.ku.dk Anders Søgaard University of Copenhagen soegaard@di.ku.dk Abstract Machine Translation operates on the premise of an interlingua which abstracts away from the surface form while preserving the mean- ing. A decade ago, the idea of using round-trip MT to guide Grammatical Error Correction was proposed as a way to abstract away from po- tential errors in surface forms (Madnani et al., 2012). At the time, it did not pan out due to the low quality of MT systems of the day. To- day much stronger MT systems are available so we re-evaluate this idea across five languages and models of various sizes. We find that for extra large models input augmentation through round-trip MT has little to no effect. For more ‘workable’ model sizes, however, it yields con- sistent improvements, sometimes bringing the performance of a base or large model up to that of a large or xl model, respectively. The round-trip translation comes at a computational cost though, so one would have to determine whether to opt for a larger model or for input augmentation on a case-by-case basis. 1 Introduction Grammatical Error Correction (GEC) is the task of detecting and correcting errors in text. It finds application in both assisted writing and second lan- guage learning. As training data for the task is scarce, efforts in this space largely focus on trans- fer learning and data augmentation. In this work, we revisit the use of round-trip Machine Transla- tion in Grammatical Error Correction, as originally proposed in Madnani et al. (2012). Machine Translation (MT) aims to preserve the meaning of text while mapping its surface form from one language into another. The ideal MT sys- tem would be robust to minor perturbations in the input text like a typo or a grammatical error, produc- ing a well-formed translation true to the intended meaning. If the translated text is then backtrans- lated into the source language, we can expect to see the original content, now free from errors. This Figure 1: Our approach. The input to the model is a concatenation of the original text and the round-trip translation. Here, English is used as target language and Romanian as pivot language for illustrative purposes; in actual experiments English is the pivot language. was the premise of the work carried out by Mad- nani et al. (2012). The statistical phrase-based MT systems of a decade ago, however, were not even close to the ideal, so the authors observed mixed re- sults in their experiments and upon further analysis concluded that the round-trip translation itself intro- duced too many new errors in the form of both un- grammaticality and loss of meaning. Modern neu- ral network-based MT systems are much stronger than their statistical predecessors. Consider the leap in BLEU score (Papineni et al., 2002) on the widely used WMT2014 …

Content truncated to 500 words