Resumen
Deep Learning has been successfully applied in hard to solve areas, such as image recognition and audioclassification. However, Deep Learning has not yet reached the same performance when employed in textual data,including Opinion Mining. In models that implement a deep architecture, Deep Learning is characterized by theautomatic feature selection step. The impact of previous data refinement in the pre-processing step before theapplication of Deep Learning is investigated to identify opinion polarity. This refinement includes the use of aclassical procedure of textual content and a popular feature selection technique. The results of the experimentsovercome the results of the current literature with the Deep Belief Network application in opinion classification.In addition to overcoming the results, their presentation is broader than the related works, considering the changeof parameter variables. We prove that combining feature selection with a basic preprocessing step, aiming toincrease data quality, might achieve promising results with Deep Belief Network implementation.