Why not use cross-entropy loss?

[TwoWhite2]

Why not use the cross-entropy loss function but only maximize the probability of a positive category？

[rhythmcao]

I suppose these two statements are the same.
If you mean why we do not directly use the CrossEntropyLoss API from pytorch for the main text-to-SQL loss, type constaints are incorporated in the decoder, such that the probability for ApplyRule/SelectTable/SelectColumn are softmaxed separately. They can be concatenated together for more efficient implementation, but directly extract the probability of a positive category is ease of use and more readable.