String-based Multinomial Naïve Bayes for Emotion Detection among Facebook Diabetes Community

Abstract This paper determined the emotions among the online diabetes community using the string-based Multinomial Naive Bayes algorithm. Facebook posts from official diabetes support groups were crawled, for a total of 15 000 pre-processed posts. Of these, 800 were manually annotated by human experts. The posts were first classified according to Plutchik’s wheel of emotions, comprising of eight dominant emotions: anger, sadness, fear, joy, surprised, trust, anticipation and disgust using the NRC Emotion Lexicon (Emolex). The emotion classifications were then refined using string-based Multinomial Naive Bayes algorithm, with results indicating a 6.3% improvement (i.e. 82% vs. 75.7% for average F-score) when compared to the Emolex-approach, and other machine learning algorithms, namely, Naive Bayes and Multinomial Naive Bayes. The higher accuracy in emotion classification reflects the feasibility of our approach. Further analysis also revealed emotions such as joy, fear and sadness to be of the highest frequencies for the diabetes community.