Online reviews are playing an increasingly important role in understanding and predicting users’rating behavior, which brings great opportunities for users and organizations to make better decisions. In recent years, rating prediction has become a research hotspot. Existing research primarily focuses on generating content representation based on context information and using the overall rating score to optimize the semantics of the content, which largely ignores aspect ratings reflecting users’ feelings about more specific attributes of a product and semantic associations among aspect ratings, words, and sentences. Cognitive theory research has shown that users evaluate and rate products following the part–whole pattern; namely, they use aspect ratings to explicitly express sentiments toward aspect attributes of products and then describe those attributes in detail through the corresponding opinion words and sentences. In this paper, we develop a deep learning-based method for understanding and predicting users’ rating behavior, which adopts the hierarchical attention mechanism to unify the explicit aspect ratings and review contents. We conducted experiments using data collected from two real-world review sites and found that our proposed approach significantly outperforms existing methods. Experiments also show that the performance advantage of the proposed approach mainly comes from the high-quality representation of review content and the effective integration of aspect ratings. A user study empirically shows that aspect ratings influence users’ perceived review helpfulness and reduce users’ cognitive effort in understanding the overall score given for a product. The research contributes to the rating behavior analysis literature and has significant practical implications.