TWO-STEP MODEL FOR EMOTION DETECTION ON TWITTER USERS: A COVID-19 CASE STUDY IN MALAYSIA

Human emotion is often a reflection of one’s behaviour that leads to awareness of one’s mental health. Emotion detection on Twitter users specifically gains attention as it generates information that is useful in the field of psychology and linguistics. However, most existing works related to the subject analysed the emotion of a tweet based on the number of keywords or phrases in it, thus resulting in false annotations and eventually false judgement of the user’s emotion. This paper proposes a two-step model with high and reliable accuracy to detect the emotion of Twitter users based on the semantic meaning of the tweets they posted. The first step classified the emotion of each tweet using four different deep learning techniques. The second step detected the emotion of the user based on the proposed statistical post-level features and a boosting technique. Then, Kappa’s agreement score method was implemented to validate both models. The first step achieved 0.9482 in accuracy while the second step achieved 0.9683 in accuracy. To further validate such highly accurate result, a case study was conducted, and a web-based system was developed to analyse the emotion of college students during the COVID-19 pandemic. Additionally, our analysis corresponded to most of the current reports on the pandemic which further proved the reliability of the developed model.