Transfer Text Of Polar Emotion Classification

Bag of Words (BOW) is now the most popular way to model text in machine learning based sentiment classification. However, the performance of BOW sometimes remains limited due to some fundamental deficiencies such as the polarity shifting problem. In this paper, we focus on polarity shifting phenomenon and conduct extensive studies on this problem. The key issues of our research are summarized as follows:First, this paper summarizes the causes and types of polarity shifting phenomenon, and compute the distribution of polarity shifting samples in the actual corpus and misclassification samples. Then we validate the polarity shifting negative impacts on sentiment classification via case study.Second, we propose a novel hybrid polarity shifting detect algorithm (HPSD), that combines the advantages of rule based methods and statistical learning based method. Then, we further propose a negation substitution method, to remove the polarity shift in negation. Based on the above two stage, a document is separated into three component parts, namely negation shift substitute part, intra-sentence polarity shift part, and unshift part. We then develop an ensemble model to train component classifiers on separated training subsets according to the types of polarity shifts. Also, a weighted combination of the component classifiers is used for prediction.At last, we propose a model, called dual sentiment classification based on artificial polarity opposite samples. We create artificial opposite samples via semantic knowledge and antonym dictionary. Thereafter, we propose two approaches called dual training (DT) and dual prediction (DP), to make use of the original samples and the created polarity-opposite samples in pairs for training a sentiment classifier and making prediction. The key issues of this algorithm is how to generate antonym dictionary, we develop two different methods to construct antonym dictionary:lexical based and corpus based.Experiments on Chinese and English standard datasets show our algorithm significantly improve sentiment classification performance, comparing with algorithm of related works.