Multilingual embeddings to scale NLP models to multiple...
summary
Shows that a prebuilt multilingual universal sentence encoder and Facebook MUSE actually align vectors across languages (MUSE for ~30 languages, Google encoder for ~16), gets normalized sentence/document embeddings you can dot to compute cosine similarity, and powers a single classifier and a semantic similarity search over mixed-language corpora. Reports practical details: English/Spanish examples, handles code-switched documents, needs GPU at scale, is cheaper than translating or training separate models and improves cross-language classification without extra labeled data.
AI-written summary · about
chapters (10)
- 0:00Multilingual embeddings intro
- 1:51Word embeddings basics
- 4:48Multilingual NLP challenges
- 6:40What are multilingual embeddings
- 8:06Facebook MUSE embeddings
- 9:10Google universal sentence encoder
- 14:15Real-world application with mixed-language documents
- 16:32Building semantic search engine
- 19:07Benefits and future of embeddings
- 20:36Q&A
related talks
41mSemantic Multimodal Multilingual Similarity Engine
23mDo Multilingual Embeddings Really Share a Semantic Space?
25mSemantic Search
31mBuilding Large-scale Localized Language Models
13mTraining Embedding Model Resiliently for Multimodal M...
42mJina Embeddings V2: From Raw Data to Bilingual Hybrid Search
36mScaling Data Beyond English: Multilingual Scaling and Data Mixing Laws
22mHow to expand your NLP Solution to new languages