Recommender systems try to discover some latent features of users and items by looking at the available information such as users' history of ratings to items and then use these latent factors to estimate users' interest level to a particular item. Traditional methods such as standard matrix factorisation rely on the ratings that users have submitted explicitly, no matter the impact of each latent factor on the total rating. Textual reviews that are posted by users can provide us some insight into the major motive behind the ratings that can also be used to explain the reasoning behind our suggestion. A first-order gradient method for matrix factorisation is proposed in this paper that considers two parameters while finding optimal latent vectors: 1) the impact of each latent factor on total rating; 2) the precision of each review. Evaluating the method on the YELP dataset shows that the algorithm converges the squared error and improves the performance remarkably.