Collaborative filtering is a widely-used recommendation technique that can provide personalized information service and thus alleviate the information overload problem. Item-based collaborative filtering algorithm serves as a cost-effective method for building recommender systems, but it still suffers from a particular kind of shilling attacks known as segment attack. The intuitive remedy is incorporating semantic information of various kinds into item similarity computation. However, extracting and syncretizing these information is often a difficult task. This paper proposes a hybrid item-based recommendation algorithm that derives the semantic correlations of items just from the information about item types by use of Bernoulli mixtures. Experimental results show that this algorithm can effectively improve both the predictive accuracy and robustness of CF systems.