In fact if you want to make any kind of multi-class classification such as sentiment analysis, you can use the exact same neutwork architecture as I presented it here. The only differente is the number of output neurons in the final classification layer. For star rating prediction you would have 5 output neurons instead of two. But as I said the pronciple stays the same. Althouh I would include maybe more neurons and layers due to higher complexity of this particular problem domain.