Spell Correction using N-Gram Modeling and Zero Shot Learning

Abstract

This paper presents an integrated spell-correction system that combines n-gram language modeling with zero-shot contextual inference using a pretrained LLaMA-2 7B model. The n-gram component efficiently generates correction candidates from local word-sequence statistics, while the zero-shot stage re-ranks those candidates by evaluating contextual and semantic plausibility without task-specific fine-tuning. The system is evaluated on two established benchmarks—BEA-60K and JFLEG—and compared against Hunspell, pyspellchecker, a standalone n-gram baseline, a standalone LLaMA-2 baseline, and the NeuSpell (BERT) toolkit. On BEA-60K, the integrated model achieves an F₁-score of 90.7%, improving over the n-gram-only baseline (59.6%) and the standalone LLaMA-2 model (80.9%). On JFLEG, the system obtains a GLEU score of 58.6, outperforming all individual baselines. An error analysis shows that the integrated model handles non-word errors with 94.1% accuracy and real-word context-sensitive errors with 85.9% accuracy. These results demonstrate that hybrid statistical–neural architectures can deliver strong correction performance while preserving the efficiency of the n-gram front end.