. 2020 Apr 13;3(2):146–150. doi: 10.1093/jamiaopen/ooaa010

Table 1.

A proposed categorization of the space of domain adaptation algorithms

Source shares	Target has	Target shares	Best methods
Labeled text	Labeled text	–	Neural feature augmentation⁵ Parameter transfer^7–9 Prior knowledge¹⁰ Instance weighting and selection¹¹^,¹²
Labeled text	Raw text	–	Neural feature correspondence learning¹⁴ Re-training embeddings¹⁹ Bootstrapping²⁰^,²¹ Adversarial training²² Auto-encoders^16–18
Labeled features	Labeled text	–	Feature augmentation⁶
Labeled features	Raw text	–	Feature correspondence learning^13–15
Trained Models	Labeled text	–	Fine-tuning²³^,²⁴ Adaptive off-the-shelf²⁵
	Labeled text	Models	–
	Raw text	–	Online self-training²¹
	Raw text	Models	Pseudo in-domain data selection²⁶

Notes: It is assumed that there is always labeled data available in the source domain. “Source shares” describes what the source site is able to share with the target site. “Target has” describes what data are available at the target site. “Target shares” describes what the target site is able to share with the source site. “Methods” gives names for the types of methods in each configuration, and citations to examples of such work