Abstract: Compared to the traditional recommendation algorithm based on item semantic with the precision of similarity calculated not high, we proposes an improved recommendation algorithm based on item semantics, through mining data sets, extracting content features, using probabilistic model and building feature vectors of items precisely, it calculates the similarity accurately between the items and combines with the results of recommendation algorithm based on semantic recommendation, not only its results is much better than the item-score recommendation algorithm for collaborative filtering and recommendation algorithm based on item semantics, but also it improves on the precision in the less users with scores.