User-generated annotations in tagging or bookmarking sites such as Flickr or Delicious can provide a promising and interesting source of information for aiding tasks such as Web resource classification. However, the use of tags brings up some challenges. Since there are no constraints on the terms that can be used for tagging, noise and ambiguity are introduced when users annotate resources. Moreover, traditional bag-of-words representations ignore connections between terms and, thus, are affected by synonymity and hyponymia. Althougth tag-based representations are a valuable source for classifying resources, the problems associated with the unsupervised nature of tags may hinder classification results. This paper presents an approach for semantically analysing social annotations in order to attain enriched concept-based representations of Web resources. Representations are enriched with concepts extracted from WordNet and Wikipedia to overcome problems caused by natural language as well as enhancing the quality of information available for performing an effective classification of resources. Several strategies for tag pre-processing, concept disambiguation and incorporation of semantic entities to representations are discussed and evaluated in this paper. Experimental results showed that the strategies proposed to associate tags with conceptual entities allow improving resource classification results, outperforming traditional approaches based on bag-of-words representations.