An out-of-distribution example has relevant conditions absent from the training or test data. A name such as Carlana or Cason is unusual because the model has not encountered it in the database. Generalization means extrapolating successfully to such examples. section 5.2.4
Unseen names
A model that predicts likely sex from known names can memorize the most common label for each name, but that strategy cannot handle original names. Character n-grams provide reusable pieces that can represent unfamiliar names. section 5.2.4
A practical approach to handling unfamiliar inputs:
- Tokenize each name into character n-grams rather than whole-word tokens. section 5.2.4
