The usage of this task type is similar as seq. However, it is not labeling the target labels directly. Instead it learns a conversion from the original word to the target label. This strategy is commonly used to convert lemmatization to a sequence labeling task. Internally, the model replace a sequence like this:
Got get
ta to
catch catch
em them
all all
! !
To a sequence like this:
Got ↓0;d¦--+e+t
ta ↓0;d¦-+o
catch ↓0;d¦
em ↓0;d+t+h¦
all ↓0;d¦
! ↓0;d¦
The labels can be used to convert the word on the left to its corresponding lemma and vice-versa. MaChAmp now performs a standard sequence labeler on these convertion-labels. During evaluation/prediction it uses the predicted convertion-labels to generate the target lemmas.
For the UD English Web Treebank for example, this reduces the label space of the lemma column from 14,909 to 268.