Trust-based recommender systems use trust relationships between users to improve the quality of recommendations. One of the most important features of trust is context-dependency. Despite the importance of context-dependency, this feature has been largely neglected in the current literature. In this paper, we propose a new approach that considers the semantic context of items to infer trust relationships between users. In this approach, the level of trust between two users varies depending on different contexts. Therefore, the trustworthy neighbors of an active user will be different for different target items, and these neighbors are determined according to the target context. The focus on context-specific ratings instead of all ratings results in fewer online computations, thus increasing the efficiency of the system as well as the accuracy of recommendations. Experimental results on a real-world data set show the higher accuracy and efficiency of the proposed approach compared to its counterparts.