I need to increase model accuracy + balance in data

I have three algorithms to sentiment classification LSTM, RNN, SVM for Arabic library and Arabic data.
The LSTM accuracy 90,
RNN, accuracy 88
SVM, accuracy 75
I need to increase the accuracy to 95 and more

The data I have is not balanced and I need to balance it.
positive 14674
negative 2953
neutral 2901

#I need to balance the following:
1 oversampling less data.
2- undersampling the most data .
3- and I have new data synthetic samples .

##I've done (smote) data, but the accuracy has dropped.