NusaBERT Teaching IndoBERT to be Multilingual and Multicultural

Published in arXiv, 2024

Indonesia’s linguistic landscape is remarkably diverse, encompassing over 700 languages and dialects, making it one of the world’s most linguistically rich nations. This diversity, coupled with the widespread practice of code-switching and the presence of low-resource regional languages, presents unique challenges for modern pre-trained language models. In response to these challenges, we developed NusaBERT, building upon IndoBERT by incorporating vocabulary expansion and leveraging a diverse multilingual corpus that includes regional languages and dialects. Through rigorous evaluation across a range of benchmarks, NusaBERT demonstrates state-of-the-art performance in tasks involving multiple languages of Indonesia, paving the way for future natural language understanding research for under-represented languages.

Recommended citation: Wongso, W., Setiawan, D. S., Limcorn, S., & Joyoadikusumo, A. (2024). NusaBERT: Teaching IndoBERT to be Multilingual and Multicultural. arXiv [Cs.CL]. Retrieved from

Many-to-Many Multilingual Translation Model for Languages of Indonesia

Published in IEEE Access, 2023

Indonesia is home to over 700 languages and most people speak their respective regional languages aside from the lingua franca. In this paper, we focus on the task of multilingual machine translation for 45 regional Indonesian languages and introduced Indo-T5 which leveraged the mT5 sequence-to-sequence language model as a baseline. Performances of bilingual and multilingual fine-tuning methods were also compared, in which we found that our models have outperformed current state-of-the-art translation models. We also investigate the use of religious texts from the Bible as an intermediate mid-resource translation domain for low-resource translation domain specialization. Our findings suggest that this two-step fine-tuning approach is highly effective in improving the quality of translations for low-resource text domains. Our results show an increase in SacreBLEU scores when evaluated on the low-resource NusaX dataset. We release our translation models for other researchers to leverage.

Recommended citation: Wongso, W., Joyoadikusumo, A., Buana, B. S., & Suhartono, D. (2023). Many-to-Many Multilingual Translation Model for Languages of Indonesia. IEEE Access.

Pre-trained transformer-based language models for Sundanese

Published in Journal of Big Data, 2022

The Sundanese language has over 32 million speakers worldwide, but the language has reaped little to no benefits from the recent advances in natural language understanding. Like other low-resource languages, the only alternative is to fine-tune existing multilingual models. In this paper, we pre-trained three monolingual Transformer-based language models on Sundanese data. When evaluated on a downstream text classification task, we found that most of our monolingual models outperformed larger multilingual models despite the smaller overall pre-training data. In the subsequent analyses, our models benefited strongly from the Sundanese pre-training corpus size and do not exhibit socially biased behavior. We released our models for other researchers and practitioners to use.

Recommended citation: Wongso, W., Lucky, H. & Suhartono, D. "Pre-trained transformer-based language models for Sundanese." J Big Data 9, 39 (2022).

Causal and Masked Language Modeling of Javanese Language using Transformer-based Architectures

Published in 2021 International Conference on Advanced Computer Science and Information Systems (ICACSIS), 2021

Most natural language understanding breakthroughs occur in popularly spoken languages, while low-resource languages are rarely examined. We pre-trained as well as compared different Transformer-based architectures on the Javanese language. They were trained on causal and masked language modeling tasks, with Javanese Wikipedia documents as corpus, and could then be fine-tuned to downstream natural language understanding tasks. To speed up pre-training, we transferred English word-embeddings, utilized gradual unfreezing of layers, and applied discriminative fine-tuning. We further fine-tuned our models to classify binary movie reviews and find that they were on par with multilingual/cross-lingual Transformers. We release our pre-trained models for others to use, in hopes of encouraging other researchers to work on low-resource languages like Javanese.

Recommended citation: W. Wongso, D. S. Setiawan and D. Suhartono, "Causal and Masked Language Modeling of Javanese Language using Transformer-based Architectures," 2021 International Conference on Advanced Computer Science and Information Systems (ICACSIS), Depok, Indonesia, 2021, pp. 1-7, doi: 10.1109/ICACSIS53237.2021.9631331.