Character Transformations for Non-Autoregressive GEC Tagging

Propose character-based method to generate target transformation instructions for GEC tagging models, as an alternative to autoregressive models. Compare character transformations to previously used word-level transformation instructions and have shown that character-based rules have better coverage and scale better in Czech, German and Russian.

本研究提出一种基于字符的方法，用于为语法错误纠正（Grammatical Error Correction）标注模型生成目标转换指令，以此作为自回归模型的替代方案。通过将字符级转换与此前使用的词级转换指令进行对比实验，结果证实，在捷克语、德语与俄语语料中，基于字符的转换规则具备更优的覆盖范围与更强的可扩展性。