In this paper we will present a new approach regarding the documents representation in order to be used in classification and/or clustering algorithms. In our new representation we will start from the classical "bag-of-words" representation but we will augment each word with its correspondent part-of-speech. Thus we will introduce a new concept called hyper-vectors where each document is represented in a hyper-space where each dimension is a different part-of-speech component. For each dimension the document is represented using the Vector Space Model (VSM). In this work we will use only five different parts of speech: noun, verb, adverb, adjective and others. In the hyper-space each dimension has a different weight. To compute the similarity between two documents we have developed a new hyper-cosine formula. Some interesting classification experiments are presented as validation cases.