When English Isn’t the Best Teacher: Source Language Effects in Cross-Lingual In-Context Learning
Summary
A broad empirical study investigates cross-lingual transfer in In-Context Learning (ICL), challenging the assumption that insights from supervised fine-tuning directly apply. The research spans seven diverse tasks, six distinct models, and a typologically varied set of languages to evaluate transfer quality. It also analyzes language confusion, a significant barrier for generative tasks within cross-lingual ICL. The findings indicate that traditional expectations derived from fine-tuning do not consistently hold true in the ICL paradigm. This suggests the need for alternative heuristics to effectively select source languages for cross-lingual ICL, moving beyond prior assumptions about data availability and linguistic similarity.
Key takeaway
For NLP Engineers designing cross-lingual In-Context Learning (ICL) systems, you should re-evaluate traditional source language selection strategies. Your prior assumptions based on supervised fine-tuning may not hold, as this study shows they don't consistently apply in ICL. Focus on developing or adopting new heuristics specifically tailored for ICL to mitigate language confusion and optimize transfer quality, rather than relying solely on linguistic similarity or data availability.
Key insights
Cross-lingual transfer in In-Context Learning (ICL) requires new source language selection heuristics, as fine-tuning assumptions do not consistently apply.
Principles
- Fine-tuning insights do not consistently apply to ICL.
- Language confusion is a key obstacle in cross-lingual ICL.
- Source language selection needs new ICL-specific heuristics.
Method
Conducted a broad empirical study across seven tasks, six models, and diverse languages, analyzing cross-lingual transfer and language confusion in ICL.
Topics
- Cross-Lingual Transfer
- In-Context Learning
- Multilingual NLP
- Source Language Selection
- Language Confusion
- Few-Shot Learning
Best for: Research Scientist, AI Engineer, Machine Learning Engineer, AI Scientist, NLP Engineer
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.