How do you handle multilingual NER when training data for minor languages is scarce?

All answers to this question.

Zero-shot transfer with XLM-RoBERTa is surprisingly effective for NER, even for languages it wasn't explicitly fine-tuned on. The model learns a universal representation of "entities." However, for better results, try "Translate-train." Translate your English NER data into Tagalog using a high-quality API, and then align the entity tags. It's not perfect, but it creates a "silver" dataset that you can use to bridge the gap. In a project I worked on in mid-2023, combining XLM-R with a small amount of translated data outperformed the zero-shot baseline by nearly 15% in F1-score.

   Answered 2025-01-13 by Shirley Jenkins

Are you concerned about "entity alignment" issues during translation? Sometimes the word order changes so much that the labels end up on the wrong words.

   Answered 2025-01-16 by Gerald Foster

• Gerald, that is the biggest headache. We use a tool called "awesome-align" which uses multilingual embeddings to find word-level correspondences between the source and translated text. It helps map the "B-PER" or "I-ORG" tags to the correct Tagalog words. It’s still a bit noisy, so a quick manual review of the top 500 rows is usually necessary to ensure the model isn't learning complete gibberish from a bad translation.

     Commented 2025-01-19 by Walter Hayes

If you have the budget, look into using LLMs to generate synthetic labeled data for the low-resource language. GPT-4 is actually quite good at generating grammatically correct Tagalog.

   Answered 2025-01-21 by Virginia Wallace

• Synthetic data is a game-changer. Using GPT-4 to create 2,000 labeled examples can give a traditional NER model enough signal to start performing quite well.

     Commented 2025-01-23 by Gary Simmons