What's in a "Domain"

 

What's in a "Domain"

Mathematically, joint distribution over inputs and outputs differs over domains 1 and 2

$P_{d1}(X,Y) \ne P_{d2}(X,Y)$

 

예를 들어,

• Content, whit is being discussed
• Style, the way in which it is being discussed
• Labeling Standards, the way thtat the same data is labeled

 

Types of Domian Shift

• Covariate Shift: The input changes but not the labeling

$P_{d1}(X) \ne P_{d2}(X)$

$P_{d1}(Y|X) = P_{d2}(Y|X)$

• Concept Shift: The conditional distribution of labels changes (e.g. different labeling standards)

$P_{d1}(Y|X) \ne P_{d2}(Y|X)$

 

Domian Adaptation

• Train on many domians, or a high-resourced domain

• Test on a low-resourced domain

• Supervised or unsupervised adaptation

 

Domian Robustness

• Train on many domains and do well on all of them

• Robustness to minority domains
• Zero-shot robustness to domains not in training data

 

출처: http://phontron.com/class/anlp2021/schedule/multitask.html