Tips are an additional information users can provide when reviewing products as a short explanation of their review. This extra informationcan be used to improve recommender systems and even provide a generated tip as an explanation of the recommendation. We look at theNeural Rating and Tips Generation (NRT) model proposed by Li et al. and attempt to reconstruct it. We then look at ways to improve thismodel; with matrix factorization, with attention, and through rating based scaled loss. We evaluate the base model and the extensionsagainst baselines with MAE and RMSE for rating prediction and with ROUGE-scores for tips generation. We also evaluate the diversityof the tips generated by measuring the frequency of bigrams. We find that we are not able to reconstruct the model by Li et al. with thesame configuration as they show, but with a different configuration, we are able to achieve comparable results. Our experiments show thatthe extension with matrix factorization named NSVDT has improved performance on rating prediction, tips generation, and diversity. Theextension with attention named NRT*A also improved on diversity and tips generation but shows a slightly worse performance on ratingprediction. The extension with rating based scaled loss named NRT*RSL shows a worse overall performance on rating prediction and tipsgeneration, but by inspecting the results for every rating individually it is clear that it has a more balanced performance across the ratings.