Ch 11 Domain Transfer with BERTApply

Carrying knowledge from a data-rich source domain into a data-poor target domain.

Core concepts

  • Domain shift. A model trained on general/source text degrades on specialized target text (legal, biomedical, product catalogs).
  • Transfer strategy. Start from a model trained on a labelled source domain and adapt it toward the unlabeled target — closing the gap without target labels.
  • Relation to the other Part 4 methods. Domain transfer is the umbrella problem; GenQ and GPL are specific solutions.

What you must master

  • Define domain shift and recognize its symptoms in retrieval metrics Level 1
  • Explain source→target transfer at a strategy level Level 1
  • Diagnose when a general embedder is under-performing on a niche domain Level 2

Architect’s lens

“Why does this generic embedding model do badly on our data?” is a question you’ll field constantly. This chapter gives you the vocabulary — domain shift — and the framing to justify a domain-adaptation workstream instead of hoping a bigger general model fixes it.