LMSpell: Spell Correction with Pre-Trained Language Models
LMSpell is a toolkit enabling spell correction using various Pre-Trained Language Models (PLMs) like encoder-based, decoder-based, and encoder-decoder models. Empirical results show that LLMs can outperform other PLM types for low-resource languages when fine-tuned with large datasets. The system includes specific methods to handle model hallucinations during evaluation.