Advanced Machine Learning to explore Public Reactions and Government Responses

• In Event: Political Methodology Virtual Poster Session

Abstract

This paper develops a hybrid approach that combines two machine learning models to improve the performance of well-known lexicons for sentiment and emotion analysis of COVID-19 twitter data. In social science, sentiment analysis plays important roles to study public opinions, negative campaigning or political polarization etc. Although standard lexicons are widely used, they do not always capture the contextual and semantic information in the short text like twitter data. This approach improves lexicon scores that characterize public sentiments (positive, negative and neutral) and emotions (fear, anger etc.). Developed method were evaluated using several benchmark twitter datasets and found effective to generate sentiment scores compared to lexicon approach alone. Finally, several sentiment timeseries were constructed using the developed approach with the COVID-19 tweets. To investigate government responses to COVID-19, two well-known events namely “lockdown or emergency declaration (LED)” and “economic support package (ESP)” were considered. Statistical testing and error analysis on the sub-timeseries around each event justified the scenario of public sentiments in each country. A change-point analysis on each timeseries also showed the distinct emotional intelligences among Japan, USA, UK, and Australia.
COVID-19 has emerged as the most aggressive disease, impacting more than 90% countries of the world. Its rapid progression has already affected human lives in many respects. Governments of various countries have already adopted several measures namely lockdown, social distancing, the declaration of economic support package etc. However, whether these responses really had impacts or to what extent they affected public lives are important research questions. Again, the quantification of major concerns or issues that are embedded in the public reactions is also important to understand people’s emotional intelligence. To address these research questions and to evaluate the algorithm, we have collected about 29000 tweets from January 2020 to June 2020 for four countries using seven keywords: “corona”, “covid”, “COVID-19”, “COVID19”, “virus”, “coronavirus”, and “new coronavirus”. These keywords were selected following the suggested lists by (Instagram, oxford dictionary, revised in 2020).
In this study, linear support vector machine and logistic regression are employed to improve the sentiment and emotion scores by VADER (Hutto et al., 2014) and NRCLex (Saif et al., 2013) lexicons. Four advanced machine approaches namely Naïve Bayes (NB), linear support vector machine (LSVM), logistic regression (LR), and random forest (RF) were first employed and cross-validated with two benchmark datasets, i.e., the “US airlines review tweets” (www.kaggle.com) and “text emotion” (medium.com) using count vector feature. Finally, two best performing classifiers, i.e., LSVM and LR were chosen for our study. Then, we developed a simple algorithm to update lexicon scores. The idea is to adjust lexicon scores by weighting for each tweet based on whether both classifiers have same/ different class decisions.
After adjusting lexicon scores by using above algorithm, we have generated two polarity (average polarity and positive-negative count ratio-pnRatio) timeseries and four emotion (fear, anger, sadness, and surprise) timeseries using our twitter dataset. Finally, two sub-series 15 days before and after each event (LED, ESP) were extracted. Statistical error analysis and Welch t-test has been performed on the subseries to identify causal relationship for each event. Experimental results showed significant changes of public sentiments in the UK and USA on the first ESP event, while the same occurred with the UK and Australian publics on the first LED event. In contrast, no significant changes in the public sentiments were observed in Japan. The verification of relevant tweets showed an increase (12%, 21%) of heath worry and anti-government sentiment tweets, while 1% decrease of economy worry tweet in Japan after the LED or ESP events. For all other countries, these parameters were decreased (USA: 8%, 20%, 8%; UK: 21%, 12%, 10%; Australia: 2%, 26%, 0%) after the LED event. Similar behavior except US were observed for the ESP event. Results also showed dominant “fear” and “sadness” emotions to Japan and USA, while dominant “anger” and “surprise” emotions to UK and USA.
Finally, we performed a change-point analysis on the sentiment timeseries using an approach based on dynamic programming (C. Trung et al., 2020). An investigation of the tweets around each dominant change points showed dominant economy worries in Japan, strong partisan-bias to USA, and strong anti-government sentiments to USA, UK, and Australia. Our results can help governments and political leaders in making policy decisions and can also be extended to develop a prediction model that can predict public sentiments in advance during similar disastrous situations.

Author

• Mdkhayrul Bashar