Online news portals produce a huge amount of content in high velocity streams. In this context, it becomes more difficult to provide dynamic real-time and large-scale recommendations that best suit each user's interests. In this article, we present a hybrid news recommendation approach based on the semantic analysis of news articles' content. The approach exploits several personalized and non-personalized approaches to alleviate the cold start problem. Experiment results in an active large-scale news delivery platform during the NEWSREEL challenge show that our system produces significantly better quality recommendations than non-semantic recommenders.