Text this: Refining the prediction of user satisfaction on chat-based AI applications with unsupervised filtering of rating text inconsistencies