Levenshtein Distance Algorithm in Javanese Character Translation Machine Based on Optical Character Recognition

Abstract

Indonesia has diverse art, cultures, and languages. Linguistically, Indonesia has many local languages, which makes it a diverse country, with Javanese being the regional language with the highest number of entries in the Kamus Besar Bahasa Indonesia. The Javanese script, one of the cultural symbols of Java, differs significantly from the Latin script commonly used in daily communication. In the context of cultural preservation, which is also one of the ministry's strategic steps, a translation or transfer process is needed from the Javanese script to the Latin script to the Indonesian language as an active participation in culture, with technology helping promote and introduce Indonesian culture. This study develops an algorithm-based approach to capture data images and improve translation accuracy. Transliteration is further enhanced by incorporating optical character recognition to convert character images. The study also applies a convolutional neural network (CNN) algorithm for character image recognition and a Levenshtein distance algorithm to translate Latin characters into Indonesian. The convolutional neural network (CNN) algorithm achieved an optimal % image detection accuracy of 95% at the 21st epoch. The translation process yielded a 90% word-level translation accuracy and 70% sentence-level accuracy. These results indicate that sentence translation remains suboptimal due to a lack of sufficient training data and similarities between scripts, highlighting the need for further improvements through transformer models or data augmentation.