Could your please give a detailed explanation for the "rank classification"?

[r-three]

Could your please give a detailed explanation for the "rank classification"?

[dptam]

1. The code listed is for training, where we do not use rank classification. Instead, we use the unlikelihood loss that tries to increase the loss for the incorrect labels (i.e. minimizing their probability). This is why there is a negative sign for cross_entropy on line 86 and we mask out the loss for the correct label on line 90.
2. It is possible to use a direct generation approach and match them against the true answer. Rank classification increases the accuracy of the model since the model only has to choose the correct choice from the list of choice, rather than having to generate the correct choice from the space of all possible outputs.