I’m trying to predict long answers from question-context pairs.
To do this, I used bigbird since the natural questions dataset does contain long answer candidates so i thought it would be a good fit. Then instead of the usual way of finding the start/end index with the highest probability directly, I set a constraint on how long the predictions should be, for example ignore all predicted spans which were too short. This way i would be able to force the prediction to be a long answer.
Would this be a reasonable way of solving the problem? Or are there any other suggested ways of doing it?