Indonesian-Kailinese Machine Translation

Abstract

The Kailinese language, spoken in Indonesia's Central Sulawesi Province, faces challenges due to limited everyday usage. According to an article published on the brin.go.id website, some dialects of Kailinese have only four family speakers. To address these concerns, this research introduces a dedicated machine translation model and dataset for Kailinese. Our translation system utilizes the IndoBART-V2 model, enabling seamless translations between Indonesian and Kailinese. We conducted two testing scenarios: one with a diverse dataset of reviews and random topics, and another focusing on review-type datasets. By employing default parameters and preprocessing techniques using the "Colloquial Indonesian Lexicon Dictionary," our translation model achieves impressive SacreBLEU scores compared to not using preprocessing or changing default parameters. For scenario 1 (Indonesian to Kailinese translation), the model achieves a score of 19.8, while for scenario 2 (Kailinese to Indonesian translation), the score is 23.0. In scenario 2, where both training and testing data consist of review-type sentences, the model achieves scores of 18.4 and 22.7 for Indonesian to Kailinese and Kailinese to Indonesian translations, respectively. These results demonstrate the effectiveness and accuracy of the developed model. Furthermore, our analysis reveals that sentence composition significantly influences the model's performance, with no notable difference between scenario 1 and scenario 2. This emphasizes the importance of considering sentence types in the translation model.