跳到正文
arXiv cs.AI· Shunta Asano, Jeonghun Baek, Toshihiko Yamasaki·· 2 小时前AI 评分33

超越双语迁移:指令微调中的多语言 Code-Switching 研究

Beyond Bilingual Transfer: Multilingual Code-Switching in Instruction Tuning

AI 导读

一项发表于 AACL-IJCNLP 2026 Findings 的研究探讨了英、日、韩、中四语言混合的指令微调数据(CSD)对大语言模型多语言能力的提升效果。在 Belebele 和 MMLU 上的实验显示,句子级多语言 CSD 在双语、三语和四语设置下均持续提高 Gemma-2-2B 的四语言平均准确率,表明多语言 code-switching 的效果不限于双语迁移。

正文 · 原文

View PDF HTML (experimental)

Abstract:Recent studies have shown that code-switching data (CSD), in which multiple languages are mixed within the same context, can improve cross-lingual transfer and multilingual alignment in large language models (LLMs). Recent studies on synthetic code-switching for LLMs have primarily focused on bilingual transfer between English and a target language, leaving instruction-tuning settings that jointly mix three or more languages within each example largely unexplored. In this work, we investigate multilingual code-switching instruction tuning across four languages: English, Japanese, Korean, and Chinese. We evaluate multilingual understanding on Belebele and MMLU. Our experiments show that simple sentence-level multilingual CSD consistently improves the four-language average accuracy across all bilingual, trilingual, and quadrilingual settings for Gemma-2-2B, indicating that multilingual code-switching can be effective beyond bilingual transfer settings.
Comments: AACL-IJCNLP 2026 Findings
Subjects: Computation and Language (cs.CL); Artificial Intelligence (cs.AI)
Cite as: arXiv:2605.29414 [cs.CL]
  (or arXiv:2605.29414v2 [cs.CL] for this version)
  https://doi.org/10.48550/arXiv.2605.29414

arXiv-issued DOI via DataCite

Journal reference: Findings of the Association for Computational Linguistics: IJCNLP-AACL 2026
Related DOI: https://doi.org/10.18653/2026.findings-aacl.2

DOI(s) linking to related resources

Submission history

From: Shunta Asano [view email]
[v1] Thu, 28 May 2026 06:03:52 UTC (4,294 KB)
[v2] Tue, 17 Nov 2026 11:28:15 UTC (4,297 KB)

来源:arXiv cs.AI · arxiv.org