<![CDATA[Spell checking is an essential component of natural language processing, as it directly influences applications such as sentiment analysis, text classification, and machine translation. Developing a reliable system for low-resource languages like Minangkabau is challenging due to frequent spelling variations and limited annotated data. This study proposes a contextual spell correction model using pre-trained IndoBERT and multilingual BERT (mBERT) embeddings applied without additional training. The method masks misspelled words, extracts the contextual embedding of the [MASK] token, and compares it with candidate embeddings generated through dictionary filtering and Levenshtein Distance. Evaluation was conducted on the Spell Error Corpus for Minangkabau Language (SPEML), which includes insertion errors, deletion errors, substitution errors, transposition errors, punctuation errors, real-word errors, and loanword errors. Results show that mBERT consistently outperformed IndoBERT, achieving an average F1-score of 0.83 compared to 0.75. Statistical validation using paired t-test and Wilcoxon signed-rank test further confirmed that the performance difference between the two models was significant. Both models reached perfect scores (1.0) in real-word and loanword categories, and strong results in insertion_medium (0.97 for mBERT and 0.95 for IndoBERT). The lowest performance occurred in deletion_short (0.52 for IndoBERT) and long words cases (0.57 for mBERT). In addition, a small-scale external validation using 100 Twitter/X sentences was conducted to assess the applicability of the proposed method to real-world social media text. Overall, the findings confirm the effectiveness of contextual embeddings for Minangkabau spelling correction while highlighting challenges in long misspelled words, deletion errors, and informal real-world text.]]>
Copyrights © 2026
