The paper discusses the prototype of a public opinion monitoring system toward Covid-19 measures using machine learning and sentiment analysis, which includes: 1) data collection; 2) data preprocessing and statistical analysis of a created corpus; 3) semantic analysis of the corpus; 4) sentiment analysis of the data. The sample of posts, which was collected for the period April-May 2020 using 360 unique search queries, includes 6726 publications of Ukrainian Facebook users. The assessment obtained by the proposed methodology is confirmed by the results of the survey about supporting the government's Covid activities, according to which only about 10% of respondents are positive about the government's actions, and more than 60% are negative. Thus, the proposed system, developed as a set of Python and SQL scripts, can be recommended for implementation.
The SARS-COV-2 pandemic has been going on already for more than a year. It is characterized by rapid spread, high mortality, and lack of specific treatment. The pandemic has become a factor that forced the governments of many countries to reconsider economic and social policies and formulate new development priorities.
The disease, which started in December 2019 in Wuhan (Hubei Province, China), was
recognized by the WHO as a pandemic on March 1
In Ukraine, the first case of the coronavirus was registered in the Chernivtsi region on
March 3, 2020; the active phase of the pandemic began on March 25
Covid-19 has an extremely negative impact on the global economy: according to the World Bank, a global GDP contraction of 5.2% was expected in 2020 (Table 1). The income per capita might fall 3.6%, leaving millions of people in poverty.
Governments are eager to find and implement proper actions to combat the coronavirus spread, as well as to overcome the negative economic consequences of the pandemic. The key to success in overcoming the pandemic is the consolidation of trust between governments and people: the effectiveness of Covid-19 control measures depends on the conscious attitude and responsibility of each citizen.
This work aims to develop a prototype of a public opinion monitoring system toward Covid-19 measures. This could be done in a form of a regular survey or by applying machine learning and sentiment analysis to the data retrieved from social media. 1 percent change from the previous year 2 estimate 3 forecast
In a pandemic, Gallup International experts have studied the issue of the government’s actions public support4. According to a three-wave survey in June 2020, the highest support5 was observed in Georgia (94%), Malaysia (94%), and the Republic of Korea (85%). The lowest levels were in Japan (34%), Bosnia and Herzegovina (35%), the United Kingdom (38%), and the United States (40%). The results in the dynamics are also quite interesting (Fig. 2): the data shows that support may increase (Kazakhstan, Malaysia, Republic of Korea, and Switzerland), decrease (Austria, Italy, Macedonia, and the United Kingdom), or fluctuate (India, Pakistan, Philippines, Russia, and the United States). 4 The question was: to what extent do you agree or disagree with the following statement – ‘I think the government is doing well with the coronavirus’ 5 Strongly agree and agree with the statement above
Unfortunately, Ukraine has not participated in this survey. To measure Ukrainians’ sentiments towards the Covid measures, we conducted an online survey6, in which 1600 respondents took part. The results showed that only about 10% of Ukrainians are positive about the government's actions, while more than 60% are negative about them (Fig. 3).
To develop a prototype of a public opinion monitoring system toward Covid-19
measures we proposed applying machine learning and sentiment analysis to the content of
the social network7. As a platform for the analysis, Facebook was chosen, because this
social network is the largest social media used by more than two billion people every
month. According to Research & Branding Group, the number of active users in Ukraine
is about 63% of the population, which is approximately 20.8 million people
The development of the system includes the following steps, which are implemented in the form of SQL and Python scripts:
I. Data collection.
II. Data preprocessing and statistical analysis of a created corpus.
III. Semantic analysis of the corpus.
IV. Sentiment analysis of the data. 6 Poll link (https://docs.google.com/forms/d/e/1FAIpQLSe3pk8AONnLr6PB1VJCBAHWzF7 tBhnS2wk8VKkEoL-SVHiwDQ/viewform) was posted on Telegram channels 7 According to the Institute of Sociology NAS of Ukraine, most Ukrainians receive information about the pandemic from the media and social networks
At the first stage, raw data was downloaded from Facebook using the DATA365 API. To identify posts discussing the Covid measures, more than 360 unique search queries have been created. The sample, which was collected for the period April-May 2020, includes 6726 publications of Ukrainian users (Fig. 4). 5). At the second stage, the texts were pre-processed: 1. Lowercase letters formatting to avoid repeating words; 2. Redundant characters removing (numbers, punctuation, etc.); 3. Stop words deleting (such as ‘this’, ‘and’, etc.); 4. Stemming (reducing words to the root form).
After the corpus had been created, the list of the most used words was analyzed (Fig.
Semantic analysis of the corpus (stage III) was performed using Word2Vec tools. The
Word2Vec model is a two-layer neural network that learns to reconstruct the linguistic
context of words. As a result, words that have a common context in the corpus are placed
close to each other in the vector space. This makes it possible to analyze the main
8 The data includes id, time, text, language, tags, reaction_count, comment_count, share_count,
reaction_like_count, reaction_love_count, reaction_haha_count, reaction_wow_count, reaction_sad_count,
reaction_angry_count
narratives. Using the Python word2vec library
CORONAVIR USstate budget epidemic complications pandemic accident outbreak measures morbidity
GOVERNME NTresistance infographics ministry continue extraordinary attenuated strengthen
As we can see from Table 2, posts about the coronavirus most often discuss the state budget and the Covid-19 control measures; in the context of the lockdown people talk about its duration, penalties, and violations, when talking about the government, it is about pandemic resistance and infographics on anti-epidemical measures.
At the fourth stage, dictionaries of positive and negative words were constructed
Analysis of fig. 7 shows that the word ‘coronavirus’ causes very controversial emotions
– ‘laughter’ and ‘anger’ are almost on the same level. The mention of the words
‘government’ and ‘lockdown’ often causes ‘anger’ and ‘sadness’; ‘mask’ – ‘love’ and
‘anger’, ‘president’ and ‘economy’ – ‘laughter’ and ‘anger’. According to Plutchik’s
wheel of emotions
These results are in line with the survey results that allow to recommend implementing the public opinion monitoring system toward Covid-19 measures based on the social network's content using machine learning and sentiment analysis.
The paper discusses the prototype of the public opinion monitoring system toward Covid19 measures based on the social network's content using machine learning and sentiment analysis, which includes: 1) Data collection; 2) Data preprocessing and statistical analysis of a created corpus; 3) Semantic analysis of the corpus; 4) Sentiment analysis of the data. The sample of posts, which was collected for the period April-May 2020 using 360 unique search queries, includes 6726 publications of Ukrainian Facebook users.
The most frequently used words of the retrieved corpus were ‘coronavirus’, ‘pandemic’, ‘lockdown’, ‘mask’, ‘power’, ‘state’, ‘president’. Semantic analysis of the corpus using Word2Vec tools showed that posts about coronavirus most often discussed the state budget and the Covid measures; in the context of the lockdown people talk about its duration, penalties and violations, when talking about the government, it is about pandemic resistance and infographics on anti-epidemical measures.
To analyze public sentiments, dictionaries of positive and negative words were created. Dictionaries comparing shows that on average, words with a positive connotation are used 30% more often than pessimistic ones. Analysis of reactions to the posts showed that the word ‘coronavirus’ causes very controversial emotions – ‘laughter’ and ‘anger’. At the same time, the mention of the words ‘government’ and ‘lockdown’ most often causes ‘anger’ and ‘sadness’, ‘president’ and ‘economy’ – ‘laughter’ and ‘anger’ (‘contempt’ and ‘aggression’ according to Plutchik's methodology).
The assessment obtained by the proposed methodology is confirmed by the results of the survey about supporting the policy responses to Covid. According to this survey, only about 10% of respondents are positive about the government's actions, and more than 60% are negative.
Thus, the proposed public opinion monitoring system toward Covid-19 measures based on the social network's content using machine learning and sentiment analysis can be recommended for implementation. 5.