Multi-Granularity Curriculum Learning for Grammatical Error Correction
A training strategy for grammatical error correction that orders and weights samples by their estimated correction difficulty at two granularities: a batch-level schedule that feeds examples from easy to hard, and an instance-level scheme that regulates the loss function per example so the model is guided toward an appropriate optimization direction. Teaches how difficulty-aware curricula improve pre-trained-language-model correction over difficulty-agnostic training.
2501.00334
This paper applies curriculum learning to Chinese Grammatical Error Correction (CGEC) - the task of detecting and correcting grammatical mistakes in Chinese sentences with pre-trained language models…