Grammatical Error Correction for Code-Switched Sentences by Learners of English

TL;DR

Code-switching complicates grammatical error correction, as monolingual GEC models fail to handle non-English constituents. The authors generate synthetic CSW data by translating spans in existing GEC corpora and reintroducing original errors, then train a multilingual GEC system using three-stage training with XLM-RoBERTa and GECToR. Evaluations on EN-ZH, EN-KO, and EN-JA with ERRANT $F_{0.5}$ show a substantial gain, especially when using linguistically informed span selection (notably translating a single noun token). The work demonstrates that data augmentation for CSW GEC yields improvements without harming monolingual performance and reveals positive cross-lingual transfer, laying groundwork for broader multilingual CSW GEC capabilities.

Abstract

Code-switching (CSW) is a common phenomenon among multilingual speakers where multiple languages are used in a single discourse or utterance. Mixed language utterances may still contain grammatical errors however, yet most existing Grammar Error Correction (GEC) systems have been trained on monolingual data and not developed with CSW in mind. In this work, we conduct the first exploration into the use of GEC systems on CSW text. Through this exploration, we propose a novel method of generating synthetic CSW GEC datasets by translating different spans of text within existing GEC corpora. We then investigate different methods of selecting these spans based on CSW ratio, switch-point factor and linguistic constraints, and identify how they affect the performance of GEC systems on CSW text. Our best model achieves an average increase of 1.57 $F_{0.5}$ across 3 CSW test sets (English-Chinese, English-Korean and English-Japanese) without affecting the model's performance on a monolingual dataset. We furthermore discovered that models trained on one CSW language generalise relatively well to other typologically similar CSW languages.