Abstarct: Contemporary search engines are not optimized to respond to users' exact inquiries. These engines simply list some relevant resources for users. Providing information through interaction is the main idea of a questions and answers (Q&A) system. Most researches which conducted on these systems focuses on the processing of questions. The processing of the answer is usually limited to choosing the appropriate answer to the previous question. In this research, polarity classification is applied to the responses of an interactive questions and answers system. Polar classification method is one of the most well-known branches in text processing. This method categorizes inputs into three positive, negative and neutral categories according to their internal content. Choosing the right features for classification is of course crucial. In this study, two new features called probability characteristics and developed point-wise mutual information have been introduced. These features, considering their tendency to one of three positive, negative and neutral categories, will set the terms. In addition to these two features, three popular features have been used in polarization are also tested. These tests are performed on unigrams and bigrams. To reduce the size of the data, information gain(GI), gain ratio (GR) and CHI square are utilized. Hitherto, polarity classification method has not been utilized on interactive texts. In a conversation, previous utterance can influence on the polarity of the answer. In this study, the effect of the previous statements has been investigated in terms of both time dependence and structural dependence. The dependence of time affects the content of the preceding statements in categorization, while structural dependence focuses on its type. The best possible precision with the unigram-baxsed feature is 77.5%. By applying the attrition, the precision becomes 75.66% and its structural dependence reaches 82.62%