Researchers at MBZUAI have developed a new automatic method to examine cross-lingual abilities in multilingual language models, testing 10 models across 16 languages. They combined beam search with language-model-based simulation, generating 6,000 bilingual question pairs and found significant performance drops compared to English, even in high-resource languages like Chinese. The method introduces perturbations to test the models' ability to transfer knowledge rather than rely on memorization. Why it matters: This research highlights critical gaps in cross-lingual AI, providing a framework for developing more equitable and effective multilingual models, especially for Arabic and other under-represented languages.
Project LITMUS explores predicting cross-lingual transfer accuracy in multilingual language models, even without test data in target languages. The goal is to estimate model performance in low-resource languages and optimize training data for desired cross-lingual performance. This research aims to identify factors influencing cross-lingual transfer, contributing to linguistically fair MMLMs. Why it matters: Improving cross-lingual transfer is vital for creating more equitable and effective multilingual AI systems, especially for languages with limited resources.
This paper explores cross-lingual transfer in Arabic language models, which are typically pretrained on Modern Standard Arabic (MSA) but expected to generalize to diverse dialects. The study uses probing on 3 NLP tasks and representational similarity analysis to assess transfer effectiveness. Results show transfer is uneven across dialects, partially linked to geographic proximity, and models trained on all dialects exhibit negative interference. Why it matters: The findings highlight challenges in cross-lingual transfer for Arabic NLP and raise questions about dialect similarity for model training.
Researchers introduce a benchmark to evaluate the factual recall and knowledge transferability of multilingual language models across 13 languages. The study reveals that language models often fail to transfer knowledge between languages, even when they possess the correct information in one language. The benchmark and evaluation framework are released to drive future research in multilingual knowledge transfer.
This paper benchmarks multilingual and monolingual LLM performance across Arabic, English, and Indic languages, examining model compression effects like pruning and quantization. Multilingual models outperform language-specific counterparts, demonstrating cross-lingual transfer. Quantization maintains accuracy while promoting efficiency, but aggressive pruning compromises performance, particularly in larger models. Why it matters: The findings highlight strategies for scalable and fair multilingual NLP, addressing hallucination and generalization errors in low-resource languages.
This paper introduces a new task: detecting propaganda techniques in code-switched text. The authors created and released a corpus of 1,030 English-Roman Urdu code-switched texts annotated with 20 propaganda techniques. Experiments show the importance of directly modeling multilinguality and using the right fine-tuning strategy for this task.
MBZUAI researchers presented a method for cross-cultural transfer learning to improve language models' understanding of diverse Arab cultures. They used in-context learning and demonstration-based reinforcement (DITTO) to transfer cultural knowledge between countries. Experiments showed up to 34% improvement in performance on cultural understanding benchmarks using only a few demonstrations. Why it matters: This research addresses the gap in cultural understanding of Arabic language models, especially for smaller Arab countries, and provides a novel transfer learning approach.
Researchers introduce PALO, a polyglot large multimodal model with visual reasoning capabilities in 10 major languages including Arabic. A semi-automated translation approach was used to adapt the multimodal instruction dataset from English to the target languages. The models are trained across three scales (1.7B, 7B and 13B parameters) and a multilingual multimodal benchmark is proposed for evaluation.