How to give input shapes to custom loss layer when batch input sentence length is different

Question

I am trying to build a Siamese network in tensorflow 2.0 for sentence similarity. It has following structure:

1. I am using the quora questions pair data. 2 input