Transfer Learning for Emotional Polarity Classification

Emotional Polarity Classification is an important task in Sentiment Analysis area. It is applied in many real problems such as reviews of consumer products and services, financial markets, and forensic analysis. The scientists from the areas of text mining and nature language processing have studied how to solve emotional polarity classification problem. They used a variety of methods, from simple methods (e.g., Lexicon-based categorization) to sophisticate methods (e.g., Statistical models). However, the problem of statistical models does not work well in a new test set whose distribution is different from training set. Therefore, the accuracy of Emotional Polarity Classification problem is still unstable. In this paper, we propose a novel approach formalism to solve this problem by using adaptation transfer learning. The transfer learning utilizes the labelled data available to solve the related but different problems. We also propose a new method that uses this approach to improve performance. The effectiveness of our approach is verified by the experiment results with two synthesis datasets and three real Twitter datasets.