Publications by authors named "Rezza Nafi Ismail"
Article Synopsis
- The -gram syllabification model faces high syllable error rates in Bahasa Indonesia due to a large number of out-of-vocabulary words, while existing models BFO and CBSPS also struggle with syllable identification and vowel detection.
- A new method, ASnGT, addresses these issues by applying syllabification at the grapheme level and eliminating reliance on vowel and diphthong detection, improving model performance significantly.
- Despite its advantages for standard words and named entities, ASnGT still has challenges in accurately distinguishing derivative words and foreign language terms.
Article Synopsis
- Recent deep learning syllabification models perform well for high-resource languages but struggle with low-resource languages like Indonesian due to limited datasets.
- The authors propose two key strategies: massive data augmentation, which includes techniques such as transposing nuclei and swapping consonant-graphemes, and a phonotactic-based validation method, to enhance model performance.
- Their findings reveal that applying data augmentation significantly boosts the dataset size and improves the model's accuracy, reducing the word error rate for both formal words and named entities in Indonesian.
