Analyzing preventive precautions to limit spread of COVID-19

With the global spread of COVID-19, the governments advised the public for adopting safety precautions to limit its spread. The virus spreads from people, contaminated places, and nozzle droplets that necessitate strict precautionary measures. Consequently, different safety precautions have been implemented to fight COVID-19 such as wearing a facemask, restriction of social gatherings, keeping 6 feet distance, etc. Despite the warnings, highlighted need for such measures, and the increasing severity of the pandemic situation, the expected number of people adopting these precautions is low. This study aims at assessing and understanding the public perception of COVID-19 safety precautions, especially the use of facemask. A unified framework of sentiment lexicon with the proposed ensemble EB-DT is devised to analyze sentiments regarding safety precautions. Extensive experiments are performed with a large dataset collected from Twitter. In addition, the factors leading to a negative perception of safety precautions are analyzed by performing topic analysis using the Latent Dirichlet allocation algorithm. The experimental results reveal that 12% of the tweets correspond to negative sentiments towards facemask precaution mainly by its discomfort. Analysis of change in peoples’ sentiment over time indicates a gradual increase in the positive sentiments regarding COVID-19 restrictions.

 Topic modeling of positive and negative is carried out using the Latent Dirichlet allocation (LDA) algorithm to analyze the factors that negatively influence the safety precautions against COVID-19.  The accuracy of the proposed framework regarding the sentiment classification is compared to a professional sentiment analysis app called 'sentiment viz app'.

Concern 3:
The theoretical framework should be discussed more in detail. Response: Thank you very much for your positive feedback and insightful comments. The proposed framework has been elaborated in the revised manuscript. Following is the newly added detail which is copied here for your convenience.

Page 10~11 of the revised manuscript
Each model in the proposed EB-DT ensemble takes the text input features to predict probabilities for positive, negative, and neutral sentiments. Mathematically it can be written as where Cp is the prediction probability by the EB-DT using soft voting criteria, N is the number of models, and avgC is the per sentiment probability by each model.
The Cp can also be written as ( (4) = ( The ensemble model EB-DT makes the final prediction using The functioning of the EB-DT can be further elaborated using an example run. For this purpose, the tweet "One of two shots so far, but have always, and will continue to wear the mask" is taken.
The average probabilities are then used by the argmax function to predict the final class of the tweet which in this case is neutral as it has the highest average probability considering the models that form the ensemble. Concerns: Thank you very much for your comments. The choice of the estimated model has been justified with respect to the existing literature. The newly added material is copied here for the convenience of the worthy reviewer.

Page 10 of the revised manuscript
The choice of an estimated ensemble model stands on several grounds. First, ensemble models have the tendency to show better performance as reported in the existing literature [32,45,55]. Utilizing more than one model potentially reduces the misclassification rate. Secondly, the choice of models varies concerning the nature of data and models can be influenced by the size of the feature set, the ratio of class samples (balanced vs imbalanced datasets), etc. which can increase the probability of model overfit or underfit. Using multiple models on the datasets and combining their output mitigates such risks. Thirdly, the voting criterion is selected because of the results reported for voting classifiers where the ensemble models tend to show better performance than individual models [33,56]. Keeping in view the diversity of the text tokens used in the tweets, it seems a probable choice to utilize multiple classifiers. Different types of machine learning models like tree-based or linear models, show a different level of performance with sentiment analysis tasks, and using more than one classifier to make the final prediction tends to increase the probability of correct class prediction.  Figure 2 and Figure 3 of the revised manuscript. A brief description is also provided on Page 7~8 of the revised manuscript.

Page 7~8 of the revised manuscript
We categorize the data into three classes positive, negative, and neutral. Among a total of 8911 records, TextBlob categorized 3193 records as positive, 4599 records as neutral while 1119 are labeled as negative, as shown in Figure 2.  Figure 3 shows the ratios of positive, negative, and neutral sentiments, as labeled by TextBlob. It can be observed that neutral sentiments make up most of the dataset with 51.61% of the dataset. It is followed by the positive and negative tweets which make up 35.83% and 12.56%, respectively. Concern 6: Diagnostic tests are absent. Response: We appreciate the reviewer's concern, however, the current study aims at analyzing the sentiments of the people using the Tweets posted on Twitter. No diagnostic tests are performed for the study.
Concern 7: Robustness checks are absent. Response: Thank you very much for your comments. For analyzing the robustness of the proposed approach, we compare its performance regarding computational complexity which is measured in terms of time required for prediction in comparison to existing works. These results are presented in Table 8 of the revised manuscript.

Page 17 of the revised manuscript
The proposed model is better in terms of computational time as compared to other studies. Although [34] provides a much smaller execution time, its accuracy is substantially low as compared to the proposed model. Also, the current study uses three models while existing studies use two models for the ensemble.