I used your Esim with bert model on my text matching data set and the performance is worce than only using BERT to finetune. So I wonder the results on your data. Does Esim with bert performs better than only using bert on your data ? Thanks