Recommendation systems help in providing suggestions regarding the products/services that the users might be interested in by analysing ratings, reviews, and other feedback obtained from the users on their previous purchases. These systems are used in various domains like e-commerce, entertainment, news, etc. In the e-commerce domain, the ratings and reviews obtained from user feedback are far less in number when compared with the number of users and items. This leads to data sparsity issues where the recommendation systems may not be able to provide accurate recommendations. To counter this issue, in our study, we used both auto-encoder and sentiment analysis on the reviews using the BERT language model to generate a more accurate version of the rating matrix. The empirical studies done on the Amazon baby product reviews dataset have shown that the approach has significantly increased the accuracy of predicted ratings.