Adversarial NLI (ANLI)
- Attempts to explicitly difficult to guess based on only the hypothesis
- Same creation as Multi-Genre NLI (MNLI), but annotators can only submit an example if a bert-style model gets it wrong
- BERT actually does pretty well as NLI, so the dataset is largely with tricky cases
Notes September 24 - 2022
- How as ANLI done recently with GPT3?