Pseudo-Error Generation for Grammatical Error Correction Based on Learner’s First Language

Anonymous

Abstract: We propose to adapt grammatical error correction (GEC) systems to the learners' first language (L1) by generating artificial errors that reflect the L1 influence. To this end, we employ two simple approaches: fine-tuning a back-translation model on L1-annotated data; and controlling the output of a back-translation model and generating artificial errors that follow the L1-dependant error type distribution. We demonstrate that, despite the simplicity of the model and the paucity of the L1-annotated data, our methods succeed in adapting GEC models to some languages. We also show that generating L1-adapted artificial errors is orthogonal to the existing method that directly adapts the GEC model to each L1. Lastly, we present an analysis of the pseudo errors generated by our models and show that they approximately capture the L1-specific error patterns.