We have acquired the data from George Washington University Libraries Dataverse, the Climate Change Tweets Ids [Data set] . This dataset has been collected from the Twitter API using Social Feed Manager, and totalled to 39,622,026 tweets related to climate change. The tweets were collected between September 21, 2017 and May 17, 2019. However, there is a gap in data collection between January 7, 2019 and April 17, 2019. The tweets with the following hashtags and keywords were scraped: climatechange, #climatechangeisreal, #actonclimate, #globalwarming, #climatechangehoax, #climatedeniers, #climatechangeisfalse, #globalwarminghoax, #climatechangenotreal, climate change, global warming, climate hoax.
Due to Twitter's Developer Policy, only the tweet IDs were shared in the database, not the full tweets. Therefore, we had to hydrate the tweet ids with the use of Hydrator application. Hydrating was carried out by us in June, 2020, and it allowed us to obtain 22,564,380 tweets (some tweets or user accounts are deleted or suspended by Twitter in its standard maintenance procedures). Challenges encountered during data hydration included dealing with deleted tweets or suspended user accounts, which is a common occurrence in Twitter's standard maintenance procedures. We addressed this by using the Hydrator application, which allowed us to recover as much data as possible within the constraints of Twitter's Developer Policy.
In order to comprehensively diagnose Polish social networks and to enable automated classification of Twitter users in terms of their attitude towards vaccinations, we collected a balanced, importance-wise database of Twitter users for manual annotation. The most important keywords used by groups that spread anti-vaccination propaganda were identified. Using our programming pipeline, databases of Polish social media on the topic of the pandemic and attitudes towards vaccinations were obtained. The raw data contained over 5 million tweets from almost 3600 users with the following hashtags related to the COVID-19 pandemic in Poland and the war in Ukraine: stopsegregacjisanitarnej, nieszczepimysie, szczepimysie, szczepienie, szczepienia, koronawirus, koronawiruswpolsce, koronawiruspolska, rozliczymysanitarystow, stopss, covid, covid19, sanitaryzm, epidemia, pandemia, plandemia, zelensky, zelenski, wojna, muremzabraunem, konfederacja, wojnanaukrainie, putin, ukraina, ukraine, rosja, russia, wolyn, bandera, upa. Twelve annotators rated the scraped Twitter users based on their posts on a nine-point Likert scale. Samples evaluated by annotators were partially overlapped in order to examine their consistency and reliability. Statistical tests performed on data before and after binning (in three- and two-category versions) confirmed significant annotator agreement. Fleiss' kappa, Randolpha, Kirchendorff alpha, and intracorrelation coefficients indicate non-random agreement among the competent judges (annotators).
Our initial data acquisition based on the abovementioned hashtags yielded 5,308,997 posts. To focus specifically on discussions related to COVID-19 and the war in Ukraine, we implemented a filtering process using Polish word stems relevant to these topics. This step reduced our dataset to 4,840,446 posts. The filtering was performed using regular expressions based on lemmatized versions of key terms. For war-related content, we used stems such as 'wojna' (war), 'inwazj' (invasion), 'ukrai' (Ukraine), and 'putin'. For COVID-related content, we used stems like 'mask' (mask), 'szczepi' (vaccine), and 'koronawirus' (coronavirus). This approach allowed us to capture various grammatical forms of these words.
Following this initial filtering, we removed three users who had no posts related to either COVID-19 or the war in Ukraine. This step left us with 3,597 users and 4,839,995 posts. Finally, to ensure consistency in our analysis, we selected only posts in the Polish language. This final step resulted in our dataset of 3,577,040 posts from 3,597 users. Before the tweets content analysis was performed, text lemmatization had been performed, special characters, links, and low-importance words based on a stop list (e.g. conjunctions) had been removed.
Data preprocessing has been carried out in Python programming language with the use of specific libraries and our original code. The hydrated tweets were further cleaned by removing duplicates and all tweets that had no English language label. Some characters and technical expressions were then replaced with natural language terms (e.g., changing “&” into “and”). We have also created a couple of versions of the database, for various purposes - in some of them we have replaced emoji pictures with their descriptions (using the demoji library and our original code), for other database versions we have removed the emojis, hyperlinks, and special characters. This caused the dataset to comprise 24,083,452 tweets (7,741,602 tweets without retweets), which makes it the biggest database of social media data referring to climate change analyzed to date.
We created the social network directed graph with the use of RAPIDS cuGraph library in Python for most of the network statistics calculations, and also with the use of the graph-tool . The final graph visualization was created with the use of Gephi after preparing and filtering the data in Python. The final graph had 4,398,368 nodes and 18,595,472 edges, after removing duplicates and self-loops.
The final label of "believer," "denier," or "neutral/unknown" was assigned to users present across annotators through the averaging of results from multiple annotators.
In the Ukraine dataset, the term 'anti-group' refers to various tactics of information warfare aimed at discrediting Ukraine's sovereignty and legitimacy, whereas the 'pro-group' consists of tweets that support Ukraine's sovereignty and legitimacy. In the Vaccine dataset, 'anti' denotes a group of users who publish tweets against vaccination, while 'pro' users advocate for vaccination programs. In the Climate Change dataset, 'denier' users dismiss it as a conspiracy theory, while 'believer' users perceive climate change as a serious threat to the future of humanity.
For ClimateChange dataset, the creationdate indicates when the connection between two users was established. The user1 and user2 fields are anonymized unique IDs representing the source and target users, respectively. Specifically, user1 is the unique ID of the source, while user2 is the unique ID of the target. The user1status denotes whether user1 is a believer (1), neutral (2), or denier (3). The creationday is a numeric value tied to the creation date. The onset and terminus fields mark the first and last days of any recorded interaction between user1 and user2, respectively, and duration captures the total time they have interacted. Finally, the w field indicates the number of interactions (such as replies, retweets, or direct messages) exchanged between them in a Twitter context.
In the Ukraine war and Vaccine dataset, the “createdate” indicates the date of that interaction. The “likecount,” “retweetcount,” “replycount,” and “quotecount” columns capture various engagement metrics on Twitter—how many times a tweet is liked, retweeted, replied to, or quoted. The “user1” and “user2” fields store unique user IDs, whereas “user1proukraine,” “user1provaccine,” “user2proukraine,” and “user2provaccine” denote each user’s stance (e.g., pro, anti, or unknown) regarding Ukraine and vaccines. The “creationday” is a numeric value corresponding to the creation date, while “onset” and “terminus” mark the first and last recorded interactions between user1 and user2, respectively. Finally, “duration” shows the total time span across which these interactions took place.
(2025)
