Indonesian Graphemic Syllabification Using n-Gram Tagger with State-Elimination

Syllabification can be approached using either grapheme or phoneme-based. Graphemic syllabification is simpler than phonemic syllabification since it does not require grapheme-to-phoneme conversion (G2P). Both phonemic and graphemic syllabification has been done on Indonesian words with average SER of 0.64% and 2.27%, respectively. The performance of Indonesian graphemic syllabification is considerably lower than the phonemic one. This research aims to improve Indonesian graphemic syllabification using a syllable boundary tagger based on the statistical n-gram model. Using fivefold cross-validation on 50k formal Indonesian words, the proposed model gives an average syllable error rate (SER) of 0.94% while the introduced state-elimination procedure reduces the SER to 0.92%, which is much lower than the previous Indonesian graphemic syllabification. Most syllabification errors come from derivative words and adapted foreign terms.