Answers are grounded in this paper and cite their sources.
This study explores the capabilities of Generative Pre-trained Transformer 4 (GPT-4), a state-of-the-art Large Language Model (LLM), in generating text in the Banjar language, an under-resourced language from South Kalimantan, Indonesia. Using prompt engineering and both zero-shot and few-shot learning approaches, GPT-4 was guided to understand and generate Banjar text. A parallel dataset and small Banjar corpus were developed for experimentation. Evaluation involved quantitative metrics (Bilingual Evaluation Understudy (BLEU) score, perplexity) and qualitative assessments from five native speakers. The few-shot approach performed best, achieving a BLEU score of 0.56 (std. dev. = $0.04)$, perplexity of 30.8 (std. dev. = 2.1), fluency of 4.2 (std. dev. $=0.3$), and coherence of 4.1 (std. dev. $=0.4$). The zero-shot approach yielded a BLEU score of 0.34 (std. dev. $=0.05$), perplexity of 45.2 (std. dev. $=3.4$), fluency of 3.2 (std. dev. $=0.4$), and coherence of 3.0 (std. dev. $=0.5$). These findings demonstrate that GPT-4, although not trained on Banjar, can generate coherent and natural-sounding text, especially with few-shot prompting. This study contributes to the development of AI technologies for local languages and highlights opportunities for integrating LLMs in cultural preservation efforts.
No takes yet. Share an insight, caveat, or question.
Syahidi et al. (2025) studied this question.