Opinion Recommendation Using Coverage for Adaptive Prediction

Abstract

Opinion recommendation aims at consistently generating a text review and a rating score that a certain user would give to a product never seen before. Inputs driving recommendation are text reviews and ratings for this product contributed by other users as well as text reviews submitted by the user under consideration for other products. The aforementioned task faces the same problems emerging in text generation using neural networks, such as repetition, specificity. In this paper, coverage loss is used as a measure of repetition in the generated text review. It is experimentally demonstrated that such a measure can be used to calibrate rating prediction and significantly improve it.