We've analysed product ratings in the past and found that the online ratings tend to have primarily bimodal distributions. People who rate online welf select and either love something​ and give very good ratings or they dislike and give poor ratings. Unlike surveys where people have to answer questions, only people with extreme reactions tend to answer these online rating questions and you don't get any meaningful distributions. So the only meaningful way to model say drivers from comments vs. ratings is to treat the ratings as binary and do a logistic type model.

Ultimately the key reasons to have these ratings in the first place is not to let people express their feelings, it is to get data for the recommendation modeling to predict better. Perhaps showing a simple like/don't like increases the number of people rating something which gives you more training data and thus better business outcome from your model without compromising on midel quality since you had to anyway model it bi-modally before.