Efficient Error-Type Transfer for Grammatical Error Detection Via Embedding Alignment

Grammatical error detection involves diverse error types, with real-word errors being particularly challenging because incorrect words remain lexically valid but contextually inappropriate. Transformer-based models achieve strong performance but typically require repeated task-specific fine-tuning, which is computationally expensive and impractical for large-scale or evolving document processing scenarios. We address this limitation by formulating cross–error-type transfer as a representation-level domain adaptation problem. Our analysis shows that fine-tuning on different error types induces systematic distributional shifts in embedding space. Instead of updating model parameters, we propose aligning target embedding distributions to a source domain, enabling zero-shot transfer while keeping the encoder fixed. We evaluate two lightweight alignment methods, PCA-based alignment and CORAL, on three linguistically complex German real-word error types (case, verb, and capitalization). Both approaches consistently outperform zero-shot baselines and recover a substantial portion of the performance gap to full target-domain fine-tuning, while reducing adaptation time and computational cost by more than an order of magnitude. The results demonstrate an efficient and practical strategy for adapting grammatical error detection systems to new error types in resource-constrained, real-world NLP applications.