Recommendation systems have emerged to select and present content according to user preferences, thus reducing the problem of information overload. Among the available techniques, the best known are collaborative and content-based filtering. Additionally, there is currently a tendency to use information provided collaboratively by users, such as tags, reviews, comments, and interactions, to reduce standard recommendation issues such as over-specialization, matchmaking, and limited content analysis. However, these annotations may contain noise, irony, and ambiguity, as well as being in a non-standard and unstructured form. Also, the semantic organization is lacking in the data so that it is possible to infer the meaning of related concepts automatically. Thus, this project aims to investigate methods of using annotations produced collaboratively by users to describe semantically the entities involved in recommendation systems. To reduce problems inherent in the use of unstructured data, we intend to develop a method that applies different techniques of feature extraction, sentiment analysis, and machine learning to obtain a rich and semantically standardized version of items and about user preferences.