Mining Emotions on Plutchik's Wheel

Abstract

Tweets embed rich information about users' moods, emotions and feelings. Mining for these latent emotions can offer clues about users' affective state on a broad range of topics ranging from their mental health to political opinions. This paper proposes a supervised machine learning approach to detect emotions from tweets. The approach is built around a Crowdflower data set of 40,000 tweets labeled with 13 distinct emotions. These 13 labels were mapped to emotions guided by the Plutchik's wheel, and are further organized into pairs of polar opposites leading to four binary classification problems: Love vs. Hate, Joy vs. Sadness, Trust vs. Disgust, and Anticipation vs. Surprise. For each classification problem, five supervised machine learning models were trained on a combination of linguistic and metadata features extracted from the tweets. The performance of these models is evaluated using sensitivity, specificity, accuracy and AUC. These results suggest that Random Forest and Support Vector Machine classifiers show close to highest accuracy in distinguishing between pair-wise emotions. Although the emotions in each pair are polar opposites on the Plutchik's wheel, their classification performance differs widely; distinguishing between Love vs. Hate and Anticipation vs Surprise show the highest (87%) and lowest (77%) accuracy respectively. Feature importance splits the discriminating power 60% - 40% over linguistic and metadata features. Our results thus suggest that every pair of polar opposite emotions are not equally differentiable, and using both linguistic and metadata features leads to better accuracy over exclusively using text-based or sentiment features.

Topics & Concepts