A Hybrid Approach to Explore Public Sentiments on COVID-19

Text processing methods like lexicon-based unsupervised approaches play important roles to quantify public opinions in the textual domain. While these have benefit directly generate sentiment scores from text data based on word-intensity scores, they perform poorly with shorter unstructured texts tweets. Besides, lexicon models often produce poor accuracy human annotated datasets. To overcome limitations of models, a new hybrid approach has been proposed. This capitalizes prediction capabilities two supervised machine learning revise using Bipolar sigmoid function that confirm better analysis. Three pre-annotated datasets used verify and proposed method. Finally, method applied corona-induced tweets, which were collected Japan, USA, UK, Australia during January–June 2020. Several emotion timeseries constructed evaluated statistical analysis against three events namely first declaration lockdown (FLD), economic support package (FEP), death-severity (FDS) event. Results showed significant reduction mean negative polarity (meanNeg) USA increase ratio between positive tweets (pnRatio) UK after FLD The people also (meanPol) FDS events, respectively. On other hand, sadness FEP, anger emotions event, surprise event shown changes. However, no emotional variables FEP any impact among four countries. Surpringly, including government responses (FLD, FEP) COVID-19 changes Japanese people. Our results can help leader’s policy decisions more accurate disaster driven sentiments.