Biomedical Machine Translation for Low-Resource Arabic-Script Languages via Cross-Lingual Transfer and LoRA Adapter Merging
Summary
This paper presents a systematic study of cross-lingual transfer for biomedical machine translation into low-resource Arabic-script languages, using Arabic and Persian as pivots. The authors evaluate LoRA adapter merging as a zero-data transfer strategy, showing it works surprisingly well for closely related languages like Dari.
View Cached Full Text
Cached at: 07/27/26, 07:40 AM
# Biomedical Machine Translation for Low-Resource Arabic-Script Languages via Cross-Lingual Transfer and LoRA Adapter Merging
Source: [https://arxiv.org/abs/2607.22300](https://arxiv.org/abs/2607.22300)
[View PDF](https://arxiv.org/pdf/2607.22300)
> Abstract:We present a systematic study of healthcare\-domain cross\-lingual transfer to address the scarcity of biomedical NMT resources for Arabic\-script languages\. We use Arabic and Persian as higher\-resource pivots to improve translation for \\textbf\{four severely low\-resource\} targets: Dari \(Afghan Persian, a standardised variety of Persian\), Pashto, Sorani Kurdish \(Central Kurdish, a major standardized variety of Kurdish\), and Urdu \(closely related to Hindi\)\. Using LoRA fine\-tuning on small decoder\-only LLMs, we train \\textit\{domain\-specific pivot adapters\} and evaluate \\textbf\{three transfer strategies\}: few\-shot in\-context learning, minimal supervised adaptation, and, to the best of our knowledge, for the first time in this setting, zero\-data LoRA adapter merging\. Supervised adaptation with just 500 sentences achieves near pivot\-language quality for Dari \(CHrF\+\+ 41\.01\) and meaningful gains for Urdu \(28\.88\), while adapter merging reaches within 3\.5 CHrF\+\+ of supervised adaptation for Dari at zero additional cost\. Pashto and Sorani Kurdish remain insufficient for high\-stakes clinical deployment exposing the limits of cross\-lingual transfer when structural distance from the pivots is too great\. LoRA adapter merging works surprisingly well for closely related languages, even without target\-language biomedical data\.
## Submission history
From: Lifeng Han Dr \[[view email](https://arxiv.org/show-email/24a5de31/2607.22300)\] **\[v1\]**Fri, 24 Jul 2026 13:44:50 UTC \(89 KB\)Similar Articles
Bridging the English-Arabic Medical Knowledge Gap: Targeted Low-Rank Adaptation via Causal Layer Selection
This paper investigates why LLMs underperform in Arabic medical tasks, showing via mechanistic analysis that knowledge exists internally but fails to surface, then proposes TLoRA, a targeted low-rank adaptation method that outperforms full-network LoRA on medical QA and introduces a new Arabic clinical dialogue benchmark.
Bridging Scientific Heritage: An Arabic--Russian Parallel Corpus and LLM Benchmark for Sustainable Knowledge Transfer
This paper presents a benchmark for Arabic-Russian scientific translation, including a hybrid parallel corpus of 27,000 sentence pairs and fine-tuned multilingual models (mT5, NLLB, Qwen) using LoRA. The best model achieves BLEU 23.15, and the work aims to lower language barriers for scientific knowledge exchange between Arabic and Russian researchers.
Cross-Lingual Transfer for Machine Translation in Turkic Languages
This paper studies cross-lingual transfer for machine translation among five Turkic languages using pairwise transfer matrices with mT5, finding that transfer is strongest between closely related pairs and that Latinization helps in script-mismatched settings.
Multilingual and Multimodal LLMs in the Wild: Building for Low-Resource Languages
This tutorial paper provides an overview of building multilingual and multimodal LLMs for low-resource languages, covering data creation, model alignment, fine-tuning, and evaluation, with a focus on practical recipes and hands-on resources.
Which Languages Transfer Best to Warlpiri? A Similarity-Based Study for Low-Resource ASR
This paper investigates cross-lingual transfer for low-resource ASR in Warlpiri, proposing a similarity-based framework combining acoustic and linguistic features to select optimal source languages. Experiments show that acoustically similar languages like Assamese and Hindi significantly reduce word and character error rates.