-> “the type of similarity” is this comment correct? , What is num_neg?, is this parameter effect training time?
Yeah, it looks like that comment should actually be for
num_neg specifics the number of incorrect intent labels that are considered alongside the correct intent label. The incorrect labels are also given similarity scores.
In terms of training time, my intuition is that increasing
num_neg should increase how long each epoch runs because you’ll be computing a larger number similarity scores. (I’d have to run an experiment to be sure though .)