Bridging Scientific Heritage: An Arabic--Russian Parallel Corpus and LLM Benchmark for Sustainable Knowledge Transfer

2026-06-29 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Intermediate, medium

Summary

A new benchmark for Arabic-Russian scientific translation has been introduced, featuring a hybrid parallel corpus of approximately 27,000 sentence pairs. This corpus, compiled from scientific abstracts and general-domain texts, aims to overcome language barriers impeding research exchange between Arabic-speaking and Russian-speaking communities, particularly for sustainability-related studies. Researchers fine-tuned three multilingual language models—mT5-base (580M parameters), NLLB-200-distilled-1.3B (1.3B), and Qwen2.5-7B-Instruct (7B)—using LoRA with various ranks. The Qwen2.5-7B model, utilizing QLoRA (rank 8), achieved the highest performance with BLEU 23.15, chrF 43.89, BERTScore 0.906, and COMET 0.758. This represents a significant improvement of +4.36 BLEU and +0.051 COMET over the zero-shot baseline. The study also found that few-shot prompting did not enhance performance, underscoring the necessity of domain-specific fine-tuning. The models, corpus, and evaluation code are publicly released to foster knowledge transfer and align with UN SDGs 9 and 17.

Key takeaway

For NLP Engineers developing scientific machine translation systems, particularly for Arabic-Russian language pairs, you should prioritize domain-specific fine-tuning over few-shot prompting. The released hybrid parallel corpus and fine-tuned models offer a robust starting point. Utilize QLoRA with a low rank, such as 8, to efficiently adapt models like Qwen2.5-7B, as this approach yielded significant performance gains (e.g., +4.36 BLEU) and is crucial for accurate knowledge transfer.

Key insights

Domain-specific fine-tuning significantly improves scientific machine translation between Arabic and Russian.

Principles

• Hybrid corpora enhance scientific translation.
• Domain-specific fine-tuning outperforms few-shot prompting.
• LoRA/QLoRA effectively adapts LLMs for specific tasks.

Method

Fine-tune multilingual LLMs (e.g., Qwen2.5-7B-Instruct) using LoRA/QLoRA on a hybrid parallel corpus of ~27,000 sentence pairs. Evaluate performance with BLEU, chrF, BERTScore, and COMET metrics.

In practice

• Utilize the released Arabic-Russian parallel corpus.
• Apply QLoRA (rank 8) for efficient LLM adaptation.
• Prioritize domain-specific fine-tuning over few-shot prompting.

Topics

• Arabic-Russian Translation
• Parallel Corpus
• LLM Benchmarking
• LoRA Fine-tuning
• Scientific Communication
• Multilingual LLMs

Code references

• sunlex0717/DissectingTensorCores

Best for: AI Engineer, Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• Ancient Greek to Modern Greek Machine Translation: A Novel Benchmark and Fine-Tuning Experiments on LLMs and NMT Models — A new AG-MG parallel corpus and benchmark significantly advance Ancient Greek to Modern Greek machine translation.
• Cross-lingual Relation Extraction with Large Language Models: Zero-Shot, Few-Shot, and Fine-Tuned Evaluation on Romanian — QLoRA fine-tuning significantly boosts cross-lingual relation extraction performance for low-resource languages, outperforming prompt-only LLMs.
• Cross-Lingual Empirical Evaluation of Large Language Models for Arabic Medical Tasks — English-centric LLMs exhibit significant performance gaps and reliability issues in Arabic medical tasks.
• IHUBERT: Vector-Based Semantic Deduplication and Domain-Balanced Pretraining for Persian Resources — Semantic deduplication and domain-balanced pretraining significantly enhance Persian PLM performance across diverse NLU tasks.
• Closing the Quality Gap in Low-Resource Text-to-Speech: LoRA Fine-Tuning of VoxCPM2 for Khmer and Korean — LoRA fine-tuning significantly improves low-resource TTS quality, but effectiveness depends on base model weakness.
• UrduMMLU: A Massive Multitask Benchmark for Urdu Language Understanding — Native, culturally-grounded benchmarks reveal significant LLM knowledge gaps in non-English, humanities-focused content.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.