After collection of possible positive/negative features with weights I used logistic regression classifier with some modifications (e.g. position algorithm) to classify the article. It determines the article polarity based on features (words, phrases and etc.)