Reproduce “Pivot Language for Low-Resource Machine Translation” 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
Abhimanyu Talwar, Julien Laasri
An investigation into using a pivot language for low-resource Neural Machine Translation demonstrated that translating Nepali to English via Hindi, using the Transfer Method, yielded a SacreBLEU score of 14.2, an improvement over the 7.6 fully-supervised direct translation baseline. The work also highlighted that backtranslation can be detrimental for highly similar language pairs like Nepali-Hindi.
Key insights
- The pivot language "Transfer Method" can lead to substantial improvements in low-resource NMT performance compared to direct fully-supervised models.
- For closely related languages like Nepali and Hindi, backtranslation may degrade translation quality rather than improve it, by introducing noise.
- Effective pivot language selection requires considering linguistic similarity and data availability for successful intermediate translation steps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.