Ch 5 Multilingual Sentence TransformersApply
Making a monolingual model understand many languages via knowledge distillation.
Core concepts
- Goal: aligned multilingual space. A sentence and its translation should map to (nearly) the same vector — enabling cross-lingual search (“query in English, retrieve German docs”).
- Knowledge distillation. A strong monolingual teacher produces target embeddings; a multilingual student learns to reproduce them for parallel (translated) sentence pairs.
- Parallel data. Training needs translation pairs rather than semantic labels — often easier to obtain at scale.
What you must master
- Explain what an aligned cross-lingual embedding space enables Level 1
- Describe the teacher–student distillation setup Level 1
- Select an appropriate pre-trained multilingual model for a language set Level 2
- Evaluate cross-lingual retrieval quality for target languages Level 2
Architect’s lens
Global products need cross-lingual retrieval without maintaining one index per language. Know the trade-off: multilingual models are convenient but often slightly weaker per-language than a dedicated monolingual model. For a client, that’s a concrete decision — one multilingual index vs. several specialized ones.