Reproduce “CANTONMT: Investigating Back-Translation and Model-Switch Mechanisms for Cantonese-English Neural 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

Kung Yin Hong, Lifeng Han, Riza Batista-Navarro, Goran Nenadic

Researchers at the University of Manchester developed NMT models and an open-source toolkit for Cantonese-to-English translation, applying back-translation and model-switch mechanisms to address its low-resource status. The best models achieved automatic evaluation scores comparable to commercial offerings like Bing and Baidu, while also releasing new parallel and monolingual Cantonese corpora.

Key insights

  • Larger NMT models, especially those pre-trained with explicit Cantonese support (like NLLB), benefited significantly from even minimal fine-tuning and generally outperformed smaller models.
  • The effectiveness of synthetic data from back-translation was highly sensitive to its quality and proportion; excessive or poor-quality synthetic data could degrade performance.
  • Integrating more real, high-quality parallel data (approximately 68K sentences) consistently yielded more substantial performance gains for larger models than relying heavily on synthetic data.

Summary by alphaXiv.

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