Opinion recommendation is a task, recently introduced, for consistently generating a text review and a rating score that a certain user would give to a certain product, which has never seen before. Input information driving recommendation is text reviews and ratings for this product contributed by other users and 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, namely repetition and specificity. In this paper, it is experi- mentally demonstrated that by employing coverage loss during training, repetition is reduced without adding extra parameters. Furthermore, the amount of repetition in the generated text review is defined as a measure of the captured information. Such measure is used to improve rating score prediction significantly during testing.
