# Language Versatilists vs. Specialists: An Empirical Revisiting on Multilingual Transfer Ability

**Authors:** Jiacheng Ye, Xijia Tao, Lingpeng Kong  
**arXiv:** [2306.06688](https://arxiv.org/abs/2306.06688)  
**Date:** 2023-06-11  
**Summary page:** https://www.summarizepaper.com/en/paper/2306.06688v1/

## Key points

- Authors Jiacheng Ye, Xijia Tao, and Lingpeng Kong explore multilingual transfer ability in English-centric models vs. multilingual pretrained models
- Research questions address the existence of multilingual transfer ability in English-centric models compared to pretrained models
- Investigation into whether this ability is dependent on English as the source language for the model
- Exploration of how multilingual transferability varies across different types of reasoning tasks
- Findings show that multilingual pretrained model does not consistently outperform English-centric model
- Observation that English may not be the optimal source language for these models, with diminishing significance as model scale increases
- Varying levels of multilingual transfer abilities observed across different task types
- Study reveals that English-centric models can possess strong multilingual transfer ability and potential to surpass pretrained models with effective training strategies

## Summary

In their paper titled "Language Versatilists vs. Specialists: An Empirical Revisiting on Multilingual Transfer Ability," authors Jiacheng Ye, Xijia Tao, and Lingpeng Kong delve into the exploration of multilingual transfer ability in English-centric models compared to multilingual pretrained models. The researchers pose several key research questions to address the gap in previous studies. They inquire about the existence of multilingual transfer ability in English-centric models and how it compares to multilingual pretrained models. Additionally, they investigate whether this ability only manifests when English serves as the source language for the model. Lastly, they explore how multilingual transferability varies across different types of reasoning tasks. Focusing specifically on multilingual reasoning ability, the authors conduct a series of comprehensive experiments across four distinct reasoning tasks to shed light on these research questions. Surprisingly, their findings reveal that the multilingual pretrained model does not consistently outperform an English-centric model. Moreover, they observe that English may not be the most suitable source language for these models and that the significance of source language choice diminishes as the scale of the English-centric model increases. Furthermore, their experiments demonstrate varying levels of multilingual transfer abilities across different task types. Overall, this study showcases that English-centric models possess not only multilingual transfer ability but also have the potential to surpass the transferability of multilingual pretrained models with effective training strategies. By highlighting both strengths and weaknesses through their experiments, the authors provide valuable insights into enhancing multilingual reasoning abilities specifically tailored for English-centric models in natural language processing applications.