Drastic increase in item/product review volume has caused serious information overload. Traditionally, product reviews are exhibited in chronological or popularity order without personalization. The review recommendation model provides users with attractive reviews and efficient consumption experience, allowing users to grasp the characteristics of items in seconds. However, the sparsity of interactions between users and reviews appears to be a major challenge. And the multi-relationship context, especially potential semantic feature in reviews, is not fully exploited. To address these problems, Multi-Task Learning model incorporating Personalized Transformer (MTL-PT) is proposed to provide users with an interesting review list. It contains three tasks: the main task models user’s preference to reviews with the proposed poly aggregator, incorporating the user-item-aware semantic feature. Two auxiliary tasks model the quality of reviews, and user-item interactions, respectively. These tasks collaboratively learn the multi-relationship among user, item, and review. The shared latent features of user/item link the three tasks together. Especially, the personalized semantic features of reviews are also fused into the tasks with the proposed personalized transformer. Two new real-world datasets for personalized review recommendation are collected and constructed. Extensive experiments are conducted on them. Compared with the state-of-the-arts, the results validate the effectiveness of our model for review recommendation.