The existing personalized recommendation algorithm for the probabilistic matrix factorization(PMF) usually ignores the description document information of the hidden items when using trust information in social networks. A PMF algorithm is proposed based on improved trust and item convolutional description information(ITC-PMF). Firstly, a new trust network is constructed by using user preferences and behavior trajectory information. Then, the item latent features are extracted from contextual documents by convolutional neural network(CNN). Furthermore, based on the PMF, the rating records, the trust information and the item description information are simultaneously used to calculate the latent feature vectors of the users and the items, so as to predict and make personalized recommendations. Finally, to verify effectiveness of the ITC-PMF algorithm, three other state-of-the-art algorithms are compared on four real-world datasets. Experimental results show that the proposed algorithm outperforms other three algorithms in terms of recommendation accuracy and robustness.