Data augmentation English-Indonesia-Madurese parallel corpus dataset using neural machine translation

Area of Science:

• Computational Linguistics
• Corpus Linguistics
• Language Technology

Background:

• Limited availability of parallel corpora for low-resource languages like Madurese hinders NLP development.
• Existing Madurese language resources are insufficient for advanced Natural Language Processing (NLP) tasks.

Purpose of the Study:

• To introduce INMAD, a novel parallel corpus of English-Indonesian-Madurese translated sentences.
• To provide a foundational dataset for Madurese NLP research, specifically at the 'engghi-enten' level.

Main Methods:

• Dataset creation involved two stages: combining source parallel corpora and data augmentation via back-translation using MarianMT.
• The corpus comprises 23,086 sentence pairs, validated by a Madurese language specialist.

Main Results:

• The INMAD dataset contains 23,086 meticulously translated sentence lines.
• The corpus focuses on the 'engghi-enten' level of the Madurese language.

Conclusions:

• INMAD serves as a valuable, validated resource for advancing Madurese language processing.
• This dataset is poised to significantly contribute to NLP research for the Madurese language.