I am wondering what could be the best approach to solve quora question similarity problem if there was no label such as “is_duplicate” available and we are forced to use a un-supervised method.I think in that case we will not be able to use SGD (backprop) to minimize loss function based on “binary_crossentropy” but what would be the training then? Similar questions should score high and dissimilar ones shouldn’t. Is there any implementation of a textual version of word2vec which can benefit from negative sampling as word2vec does but on textual data? Is there any other method that I should be thinking about? I am already looking at Siamese LSTM as unsupervised manner but not sure how to train where we don’t have a label. Also, should I take any different approach?