Comparison of mT5 and XLM-R Inference Latency in Zero-Shot Cross-Lingual NER Across 100 Low-Resource Languages

Abstract: Despite the emergence of large-scale multilingual pre-trained models like mBERT, XLM-RoBERTa, and mT5, natural language processing (NLP) still struggles in low-resource languages due to limited annotated data. This paper explores the use of transfer learning to adapt pre-trained multilingual models to low-resource tasks such as Named Entity Recognition (NER), sentiment analysis, and machine translation for languages like Amharic, Hausa, and Sinhala. By leveraging zero-shot and few-shot learning paradigms and evaluating cross-lingual embeddings and token overlap, we demonstrate signif Research goal: How does the inference latency of mT5 compare to XLM-R on zero-shot cross-lingual NER tasks when scaled to 100 low-resource languages, measured using throughput (queries per second)? Autonomous synthesis report generated by Assignee Research. Tribunal consensus score: 8.0/10. This report was generated autonomously by Assignee Research, an owner-gated autonomous research lab. The content synthesizes findings from peer-reviewed papers. Tribunal score: 8.0/10.