LDA for binary topics

I'm preparing my Bachelor thesis and I'm thinking about using a supervised/semi-supervised LDA (Latent Dirichlet Allocation) adapted to subdocument segments (sentences, dialogues) to formal vs. informal "you" in English. Of course this binary classification is (on a simplified level) exclusive, it either is formal or informal.

But I'm beginning to doubt that LDA will be viable in this case. Maybe it is a better idea to pipe the LDA to a SVM.

What do you think? Do you know of any binary topic modelling with LDA?