Abstract
Text emotion classification is a challenging topic in the Text Mining field. Quantifying textual information can be done with various approaches using words, character n-grams or patterns. In this research, we will explore and highlight some of the interesting properties of using text-based patterns for emotion classification. Emotions like joy and sadness are often used and clearly expressed on social media; whereas, emotions such as fear or disgust are more sparse and less abundant. This unbalanced data makes the performance of the classifier inconsistent over the different emotions. In addition to this unbalanced emotion challenge, text data on micro-blog, comments, and short messages are considered quickly composed spoken language text. Social Science and Psychology researchers gave definition about emotion and describe similarities and distances between the different emotions. For instance, anger is closer to disgust than it is to joy. This paper describes an approach to use this prior knowledge in order to improve short text emotion labeling. This prior knowledge about emotion distances and similarities is used to transfer text feature learning on an emotion to other emotions. This transfer knowledge approach ensembles with another emotion classifier improve the performance of this emotion classifier. We use ranking and multi-label metrics to compare different models. Our experiments show that classification scores for rare emotions as well as multi-label and ranking performances have increased.