How to expand your NLP Solution to new languages
summary
A CV phrase-extraction system uses cross-lingual embeddings (trained with word2vec on domain data and aligned via CCA) plus a BiLSTM+CRF sequence-labeler and achieves about 75% F1 zero-shot on German without German labels. Adding a few hundred target examples gives a small boost (about 4% lift with 200 target examples). Key practical findings: build bilingual lexica from domain data of about 5k top frequent domain words, and prefer CCA alignment with word2vec over generic pre-trained vectors.
AI-written summary · about
chapters (13)
- 0:00Multilingual NLP expansion
- 2:48Multilingual support challenges
- 3:52Cross-lingual embeddings
- 5:41CCA for shared spaces
- 7:46Zero-shot and joint training
- 8:28Transfer learning validation
- 11:16Bilingual dictionaries
- 13:26Dictionary sources impact
- 14:32Dictionary size and frequency
- 16:15Visualizing embeddings
- 18:43Error analysis
- 20:29Results across languages
- 21:33Key takeaways
related talks
33mScalably Internationalizing Millions of Latent Semantic Labels
45mHow multilingual is your NLP model?
31mRemaining Challenges in Deep Learning based NLP
22mMultilingual embeddings to scale NLP models to multiple...
42mJina Embeddings V2: From Raw Data to Bilingual Hybrid Search
33mBulk Labelling Techniques
31mTaking AI/ML to the next level: Text!
31mKeynote Q&A – Barbara Plank