Ch 6 Unsupervised Training (TSDAE)Apply
Building sentence embedders when you have text but no labels — the common real-world situation.
Core concepts
- No labels, still learn. Unsupervised methods learn embeddings from raw text alone.
- TSDAE (denoising auto-encoder). Corrupt a sentence (delete/swap words), encode it into a single vector, then force a decoder to reconstruct the original — the bottleneck vector must capture meaning.
- Use as a starting point. Unsupervised pre-training followed by light supervised fine-tuning is a powerful, data-efficient combo.
What you must master
- Explain when unsupervised training is the right choice (no labels available) Level 1
- Describe the TSDAE corrupt–encode–reconstruct loop Level 1
- Run unsupervised pre-training on a raw domain corpus Level 2
- Combine unsupervised pre-training with later supervised fine-tuning Level 2
Architect’s lens
Most enterprises have plentiful unlabelled text and almost no labelled pairs. This chapter unlocks the “cold start” strategy: bootstrap a domain embedder from raw documents, then refine. It sets up Part 4’s more advanced domain-adaptation techniques.