Perceptron is just a simple binary classifier, that works on fixed size vectors from R^n as input data. So in order to use it you have to encode each of your documents in such a real-valued vector. It could be for example a bag-of-words representation (where each dimension corresponds to one wor, and the value to number of occurences), or any "more complex" representation (one of which is described in the attached paper).
So in order to "port" perceptron to sentiment analysis, you have to figure out some function f, that feeded with document returns real-valued vector, and then train you perceptron on pairs
(f(x),0) for negative reviews
(f(x),1) for positive reviews