We present a description of our system submitted to the VLSP 2016 Evaluation Campaign of Sentiment Analysis for the Vietnamese Language. This year the campaign focussed on polarity classification, i.e., to classify Vietnamese reviews or documents into positive, negative or neutral. In order to address the task, we implemented a multi-layer neural network-based method that uses three types of features as input. Our internal evaluations indicate that by using TF-IDF feature to represent sentences, we can obtain the best performance with 66% of precision and 65% of recall. The official accuracy of the proposed method on the testing set (evaluated by the organiser) is 65.9%.