lexical note hlt · adv-nlp

adv-nlp-8

Interpretability

Challenge set

Test sets are often sourced from the same dataset as the training data. Challange sets aren’t. Challenge sets can include negative examples and ill-formed sentences. Challenge sets are systematically crafted to target one or more linguistic capabilities.

Model independent. Can be used for any model you want.

Adversarial examples

Are based on the model and challenge the model’s robustness.

https://bibinlp.umiacs.umd.edu/sharedtask.html