Abstract
Recent advances in artificial intelligence have significantly improved automatic translation, yet historical and extremely low-resource languages remain challenging due to limited training data, complex morphology, and orthographic variation. This study investigates the effectiveness of modern neural translation approaches for translating Chagatai historical texts. The written heritage of Central Asia contains many historical documents composed in the Chagatai language, which served as a major literary and administrative language across the region from the Timurid period until the early twentieth century. Although many manuscripts have been digitized, the majority remain inaccessible to broader scholarly and public audiences because reliable translations into modern languages are scarce. In this study we constructed an experimental parallel dataset by web scraping and custom extraction scripts. Dataset contains 10712 aligned entries in Chagatai paired with English and Kazakh translations. Preprocessing steps includes Unicode normalization, script standardization, and multilingual alignment. After that we use fine-tuning for two different translation architectures: a multilingual NLLB-200 neural machine translation model trained within the No Language Left Behind framework and a TranslateGemma generative large language model instruction-tuned for translation tasks. For the NLLB-200, the tokenizer was extended with a chg_arab token, and training used a Seq2Seq framework with AdamW optimization, mixed precision, and early stopping based on BLEU scores. For TranslateGemma, parameter-efficient fine-tuning with DoRA adapted the model’s key projection and embedding layers to Chagatai orthography, leveraging typological similarity with Uzbek language, while regularization and precision optimizations ensured stable convergence on the limited dataset. The results show that the NLLB-200 multilingual model consistently produces more reliable translations, achieving higher scores in most evaluation metrics and demonstrating stronger semantic fidelity, particularly when translating into the related Turkic language Kazakh. In contrast, the generative model often produces fluent but less faithful translations and exhibits a higher rate of hallucinated content. For example, XCOMET showed that NLLB-200 consistently outperformed TranslateGemma, achieving 74.97 for English and 74.31 for Kazakh, compared to Gemma’s 73.80 and 69.75, respectively. These findings suggest that source-anchored multilingual translation models, such as NLLB-200, are better suited for translating extremely low-resource historical languages like Chagatai.