ABSTRACT
Social media especially the Twitter platform has become a good data-source in Japan for tracking various social issues including depression and other mental health problems. It can overcome the under-representation and sampling bias of the survey data. In this study, we develop a machine learning approach to predict depression of Japanese people and compare their depression levels between pre-pandemic (2018) and pandemic (2020) times. We use three datasets in this study in which the first dataset is used for model development and its validation, while the rest two are used as test datasets for depression prediction. These two datasets represent timeseries tweets for the years 2018 (pre-pandemic) and 2020 (pandemic), respectively. After preprocessing the tweets, the Bag-of-words (BOW) feature is computed for each test dataset, which is later fed to the trained Logistic Regression (LOGR) model to classify tweets into "Depressive"and "Non-Depressive"categories. An analysis on the classified tweets shows a significant increase of depressive tweets in 2020, when compared with those in 2018. The covid related depressive tweets was found 50.37% of the total covid-related tweets and 8.6% of the total depressive tweets in the 2020 dataset, which indicates an increased impact of depression on the Japanese people due to COVID-19. Also, the peak depression occurs in June and August 2020 just after the first peak of the death progression timeseries in Japan, which indicates the consequences or shocks of exponential death-turmoil along with the increasing economic uncertainty and mobility restrictions. The timely application of our method to suitable textual datasets can minimize the calamity of future disasters like COVID-19 as well as it can help making suitable policy decisions for sustainable solutions against depression. © 2022 IEEE.
ABSTRACT
Text processing methods like lexicon-based unsupervised approaches play important roles to quantify public opinions in the textual domain. While these methods have benefit to directly generate sentiment scores from text data based on the word-intensity scores, they perform poorly with shorter unstructured texts like tweets. Besides, these lexicon models often produce poor accuracy with the human annotated datasets. To overcome these limitations of lexicon models, a new hybrid approach has been proposed. This new approach capitalizes the prediction capabilities of two supervised machine learning models to revise the lexicon scores using Bipolar sigmoid function that confirm better accuracy in the sentiment analysis. Three pre-annotated datasets have been used to verify lexicon-based models and the proposed hybrid method. Finally, the proposed method has been applied to the corona-induced tweets, which were collected from Japan, USA, UK, and Australia during January–June 2020. Several sentiment and emotion timeseries have been constructed and evaluated using statistical analysis against three events namely the first declaration of lockdown (FLD), the first declaration of the economic support package (FEP), and the first death-severity (FDS) event. Results showed the significant reduction of the mean negative polarity (meanNeg) in the USA and the significant increase of the ratio between positive and negative tweets (pnRatio) in the UK after the FLD event. The UK people also showed significant reduction of the mean polarity (meanPol) after the FLD and FDS events, respectively. On the other hand, the sadness emotion in the UK after the FEP, the anger and sadness emotions in Australia after the FDS event, and the surprise emotion in the UK after the FDS event have shown significant changes. However, no emotional variables after the FLD event and no sentiment variables after the FEP event have shown any impact among the people in any of the four countries. Surpringly, no events including government responses (FLD, FEP) to COVID-19 showed significant changes to the emotions of Japanese people. Our results can help leader’s policy decisions and can also perform more accurate prediction of the disaster driven public sentiments. © 2022, The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd.
ABSTRACT
The COVID-19 has affected human lives in many ways throughout the globe. Several recent studies indicated that it has greatly impacted people's income. Many people have become jobless and many business entities have already closed especially in the travelling, tourism, entertainment, and restaurant sectors. Anecdotal evidences suggest that the pandemic has severely affected mental health issues. However, systematic studies for tracking public worries towards health and economy due COVID-19, regarded as hWorry and eWorry, over time are still lacking. In this study, several supervised machine learning models have been applied to a collection of public tweets to explore the mentioned worries. Experimental analysis with a set of 4072 tweets spanning six-month from January 2020 to June 2020 have discriminated tweets into hWorry and eWorry classes with 61% accuracy. By applying a lexicon-based approach to the classified tweets, our approach reveals how the four selected emotions, i.e., fear, anger, sadness, and trust propagated over time. While the fear and trust emotions showed dominant temporal patterns in both classes, the average anger and sadness emotions were stronger in the e Worry class as compared to those in the h Worry class suggesting the necessity of more viable economic policies to overcome corona calamity. © 2021 IEEE.