Siamese Networks For Comparison

Can we use siamese networks for other comparison purposes like comparing two tweets to see which one is more positive or negative?

Siamese networks are indeed used for comparison, but generally to see whether two input samples are similar or dissimilar, so in a classification way. To know which of the two input tweets is more negative or positive would require having a latent space that describes this metric and a way to include the sign of the distance (dist(A, B) and dist(B, A) should be different by a sign). Theoretically, if you create a regressor you would be able to do that, but I think having classes and structuring them properly would be more accurate unless there is a real need for that. But even in that case, I would probably try to do a proxy problem using classification if possible.