Reproduce “Frustratingly Easy Label Projection for Cross-lingual Transfer” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yang Chen, Chao Jiang, Alan Ritter, Wei Xu

EasyProject, an optimized "mark-then-translate" technique for cross-lingual label projection, achieves improved performance for span-level NLP tasks across 57 languages. This method surpasses traditional alignment-based approaches by utilizing square bracket markers and efficiently fine-tuning multilingual machine translation models for robust marker preservation.

Key insights

  • A simple, optimized marker-based projection (EasyProject) can consistently outperform more complex, pipeline-driven alignment-based methods for cross-lingual label transfer.
  • Efficient, targeted fine-tuning of large multilingual MT models can significantly improve their ability to preserve structural markers, even with minimal computational cost.
  • Accurate preservation of span boundaries is a critical factor for successful cross-lingual transfer in span-level NLP tasks, where EasyProject demonstrates superior performance.
  • The approach yields particularly substantial gains for low-resource languages and those with non-whitespace scripts, highlighting its utility in challenging linguistic contexts.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.