Introduction

Profound changes in the information society have led to a sharp increase in the volume of data available on the Internet. Since 2010, digital content on the Web has grown exponentially, rising from 2 Zetabytes in 2010 to approximately 97 Zetabytes in 2022 [1], with a forecast of exceeding 180 Zetabytes in 2025. This indicates a trend of doubling the amount of online data over the next two years.

The widespread use of computers—including desktops, notebooks, and, especially, smartphones—has not only facilitated the consumption but also the massive production of digital content by internet users, particularly through interactions on social media platforms.

The term “big data” emerges in this context to define the massive amounts of digital data produced. Tsai et al. [2] identify speed, volume, and variety as the main components of big data. In this study, big data alludes to various facets of the process of extracting informational value and sheds light on the challenges of identification, collection, storage, analysis, and visualization of large volumes of data.

Among big data sources, social media platforms such as Facebook, Instagram, and Twitter (renamed as “X” in July 2023) stand out. Users create accounts or profiles on these platforms, share information, and connect with others by following their respective accounts.

The fundamental conception underlying social network theory is that apparently autonomous individuals and organizations are, in reality, embedded in social relationships and interactions [3].

Pew Research reports have extensively documented how the emergence of social media platforms has influenced communication patterns worldwide, work and consumption behaviors, communities, and how individuals obtain and share information about health, politics, civic life, relationships, and even people’s stress levels [4].

In politics, social media has proven to be an essential tools in electoral campaigns, aiding in building public image and support networks for politicians. Simultaneously, it is utilized to deconstruct opponents’ images through negative campaigns, dissemination of fake news, and hate messages, including the use of bots.

Social media platforms played a prominent role in significant recent events, including the 2016 US presidential campaign of Republican candidate Donald Trump, the Brexit referendum in the same year leading the UK to initiate the process of leaving the EU, and Jair Bolsonaro’s 2018 presidential campaign in Brazil.

Therefore, data from these platforms have garnered the attention of researchers from various fields of knowledge. Social media data analysis can offer insights into human behavior and interaction, contributing to a deeper understanding of public opinion on specific topics. It aids in identifying population niches, studying group changes over time, pinpointing influential social agents, and even developing strategies for product or service recommendations [5].

The process of gathering this data is known as data scraping, which involves searching for hidden information within a large dataset using algorithms [6]. In addition to data scraping, data mining and information retrieval are employed to manage, process, analyze, and visualize the extensive amount of structured or unstructured big data. This field has experienced exponential growth and has become increasingly institutionalized in the 21st century [7].

In this context, the application programming interface (API) technique has been widely employed to extract large amounts of data from social media platforms (referred to as big social data) [8–10], as was the case in this research.

However, it is worth noting that the structure of APIs used by major social media platforms like Twitter (X) and Facebook imposes significant restrictions on research possibilities, tool usage, and the types of data that can be collected [11].

The main objective of this study is to describe the process of collecting data extracted from Twitter (X) during the Brazilian presidential elections in 2022, encompassing the post-election period and the event of the attack on the buildings of the executive, legislative, and judiciary branches in January 2023. Additionally, the study provides an overview of the general characteristics of the dataset created from 282 million tweets, named “The Interfaces Twitter Elections Dataset” (ITED-Br). The acronym ITED-Br alludes to the research group responsible for its creation, namely, “Interfaces—Center for Sociopolitical Studies of Algorithms and Artificial Intelligence.

This article is structured as follows:

1. The literature review section identifies available datasets and corpora containing data from discussions regarding political events on Twitter (X). It reviews articles that work with datasets featuring the most extensive corpus of political data, positioning our dataset within this literature;
2. In the materials and methods section, the challenges encountered in obtaining data based on the use of the Twitter (X) platform API are detailed, along with the strategies adopted in the face of the imposed restrictions. Additionally, this section presents the study flowchart;
3. Subsequently, the results of the study are presented, featuring an exploratory analysis of the corpus, descriptive measures of the dataset, and sample analyses;
4. The next section presents the discussion of results;
5. The Conclusion addresses the challenges of creating a database of this magnitude, especially in the context of collecting, storing, and using big social data. It also offers suggestions for future research and potential uses of the database.

The main contributions of this article consist of:

• Describing the challenge of collecting, storing, and utilizing big social data, particularly in the context of scientific research.;
• Presenting and analyzing data on a relevant topic, specifically the use of social media platforms in one of Brazil’s most polarized electoral campaigns and one of the most polarized globally. The second round of this election featured Luiz Inácio Lula da Silva (Workers Party—PT, left-wing) and Jair Messias Bolsonaro (Liberal Party—PL, far-right wing);
• Providing relevant data for studies on sociopolitical phenomena and dynamics within social media platforms;
• Making the third-largest corpus of political data from social media available for areas that require large volumes of real or informationally-rich data, which does not always apply to synthetic datasets. The database will be made available on GitHub.

Making it available means offering access as conveniently as possible without losing data or important data characteristics.

This work was motivated by the substantial amount of data often existing in a ‘public’ access environment regarding social processes or dynamics. However, these data are not necessarily available in a form that can be easily understood or from which value can be extracted by those who could, and often should, benefit from such information.

Literature review

The literature review aimed to identify previously published datasets and corpora containing data from discussions about political events on Twitter, along with articles featuring extensive big social data corpora found in the literature.

The search was conducted in the main collection of the Web of Science database, considering all publications—scientific articles, book chapters, and works presented at events—published between January 1, 2010, and June 1, 2023. The search focused on studies about the use of Twitter (X) for political purposes, analyzing messages extracted from the platform, commonly referred to as “tweets.” The keywords “Twitter” must appear in the text, along with the terms “politics” or “election” and “corpus” or “dataset.”

The identified studies had diverse objectives, including analyzing the population’s perception of political events, examining the use of Twitter by political figures and/or parties, detecting bots, fake profiles, or spam on Twitter, identifying misinformation or hate speech in tweets, predicting the results of an election, analyzing relationships between traditional media and Twitter, proposing a method or model for analyzing or collecting tweets and presenting and/or discussing a dataset/corpus. However, this review will focus specifically on works featuring an extensive corpus and articles to offer a detailed dataset, making it publicly available for future research.

Thus, only 13 articles out of the 153 found in the search met the criteria of having an extensive corpus and a detailed dataset and were consequently selected.

Regarding the language of the available datasets, 6 have tweets in English; 1 in Spanish; 1 in Italian; 1 in German; 1 in Arabic; 1 in English and Hindi; 1 in English and Pidgin, and 1 dataset was multilingual.

Table 1 displays the 13 corpora from the selected articles in this literature review, with the dataset built in this study positioned among them in order of size:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Corpora from the articles in order of size.

https://doi.org/10.1371/journal.pone.0316626.t001

The largest corpus was published by Chen, Deb, and Ferrara [12], and it is the first public dataset with tweets from the 2020 United States presidential elections, available on GitHub. The collection of tweets began in May 2019 and covers the Republican and Democratic primaries; the confirmation of Donald Trump and Joe Biden as candidates for their respective parties; the controversy regarding the postal vote system introduced as a result of the COVID-19 pandemic; confirmation of Joe Biden’s victory; the recurring allegations made by Donald Trump that the election had been rigged and the other events that culminated in the invasion of the Capitol on January 6, 2021. As of January 22, 2021, the dataset had 1,258,209,617 tweets in several languages, and the authors intended for the collection to extend to the first six months of Joe Biden’s administration.

As this dataset focuses on the US elections, the predominant language of the tweets is English. The dataset also includes data such as the tweet’s publication date, information about the author, and whether the tweet was original or a reply, retweet, or quote to another tweet. In addition to the data on tweets and users, the authors observed that less than 1% of tweets have information about the author’s location and developed their own technique for identifying location, described in the article. Finally, one of the limitations of the dataset is the collection format allowed when the work began. At that time, the Twitter API permitted access to approximately 1% of the flow of all tweets in real-time and returned those that had any of the keywords pre-established. However, in 2021, Twitter launched a license for academic research, which allows researchers full access to the files found in the search.

The second-largest corpus was presented by the research of Kandasamy et al. [13]. Faced with one of the challenges posed by the COVID-19 pandemic, they collected 312 million tweets in English about the pandemic and proposed a method for analyzing sentiments. The authors were based on the enormous flow of information on social media, accompanied by an erroneous orientation that resulted in complications for the health sectors of governments in several countries. The research method employs the N-gram autoencoder integrated into a machine learning architecture. Using four classification algorithms—Decision Tree, Support Vector Machine (SVM), Random Forest, and K-Nearest Neighbors (KNN)—the accuracy of the sentiment analysis approach was 87.75%.

The literature review provides a framework to position the dataset presented in this study among similar ones. This dataset, named ITED-Br—Interfaces Twitter Elections Dataset 2022, gathered tweets posted during the 2022 presidential elections in Brazil. When considered alongside the 13 datasets identified in the literature review, ITED-BR ranked as the third most extensive dataset of tweets with political purposes. It comprises more than 280 million tweets (precisely 282,135,572) and can be regarded as the most comprehensive when evaluating datasets in languages other than English.

With the fourth largest corpus, Bovet and Makse [14] analyzed the dynamics and influence of fake news during the 2016 United States elections. In their research, the authors used a dataset containing 171 million tweets [15] collected in the five months prior to election day and identified 30 million tweets made by 2.2 million users that contained links to news outlets. As a result, 25% of these tweets spread false or biased news, and Donald Trump’s supporters were responsible for influencing the dynamics of fake news spread, even though the center and left-wing news disseminators were the most influential on the platform.

Panda et al. [16] had the fifth largest corpus of the review, with around 120 million tweets. The authors’ dataset, PoliTwictionary, included around 80 million tweets from Indian politicians and 40 million tweets from American politicians. The primary objective of the study was to introduce the dataset formed of tweets from political actors from India and the United States built by NivaDuck, a two-step classification pipeline designed by the authors. NivaDuck was used to identify politicians on Twitter through the description section in their accounts and the content of their tweets. According to the researchers, it identified more than 18,500 Indian politicians and more than 8,000 American politicians, and the dataset was complemented by human verification.

Sixth, with 107 million tweets from 246,000 users, Fink et al. [17] constructed a dataset to analyze the sentiments of Nigerian Twitter users during the 2011 elections and compare the results obtained on Twitter with the results presented by electoral polls. The authors’ collection spanned between April 2010 and April 2011, using Twitter’s API 1.0 geographic query feature. When comparing, the authors concluded that the correlation between sentiment analysis and electoral polls was not significant and that Twitter, due to the freely discussed topics, should be used as a complement to opinion polls and not as an instrument for assessing public opinion.

The seventh largest corpus was introduced by Cardaioli et al. [18], offering a dataset consisting of 6,685 users and 9,593,055 Italian tweets that were labeled according to political orientation. The collection was carried out in two stages. The first started by identifying a thousand users who expressed their political orientations through retweets or comments on party publications, and the second was the random selection of Italian Twitter users. In addition, the authors discuss the feasibility of automatically classifying political orientation using machine learning techniques.

In the article “Twitter social bots: The 2019 Spanish general election,” Pastor-Galindo et al. [19] present the eighth-largest dataset comprising 5.8 million tweets related to the 2019 elections in Spain. The authors detail the collection carried out between October 4 and November 11 using 46 hashtags and argue that one of the objectives of the work is to enable researchers from different countries to improve the stages of data collection, organization, and pre-processing, as well as processes presented in detail in the article. The work also discusses the detection, analysis, and classification of social bots on Twitter.

The ninth largest dataset was called “EPINetz Twitter Politicians,” published by König et al. [20]. It gathers 426,614 tweets in German written by 2,449 German parliamentarians, ministers, state secretaries, parties, and ministries throughout 2021. The authors carried out an exploratory analysis to demonstrate the applicability of the dataset during the 2021 German federal elections, in addition to expressing the intention to update the corpus annually for possible longitudinal analyses.

The work of Thorburn, Torregrosa, and Panizo [21] introduces the Alt-Right Twitter Accounts dataset, ranking as the tenth largest. It comprises 422 Twitter users associated with the “Unite the Right” rally, an extremist movement originating in Charlottesville, United States, in 2017. The dataset contains 123,295 tweets, and the authors believe that these messages will contribute to studies on radicalization in online environments. Additionally, they aim to enhance understanding of the language used by the movement members.

Green and Larasati [22] published the eleventh dataset analyzed in this study, called “The First 100 Days,” containing tweets in English. The collection is based on tweets made by the President of the United States and Senators in the first 100 days of his term, resulting in a corpus of 59,789 tweets. The dataset aims to identify the process of imposing agendas by parties on the new president and assist future work with a focus on linguistics and Natural Language Processing (NLP), emphasizing the relationship between language and politics.

Ghosh et al. [23] presented the twelfth dataset, composed of more than 10,000 tweets in English, called “Multi-domain hate speech corpus” (MHC). It contains hate speech messages against religion, nationality, ethnicity, and gender. The corpus, collected between January 2018 and May 2020, was manually annotated to distinguish between tweets containing or not containing hate.

Jaidka [24] published the thirteenth largest dataset, the “Twitter Deliberative Politics dataset,” featuring 6,000 political tweets labeled according to their deliberative characteristics. The collection was carried out between January 2017 and March 2018, using a sample of 1% of the responses made by Twitter users to the 536 congressmen in office in the United States.

The last dataset on the list of the largest found in the literature (14th presented in Table 1) was published by Alzahrani et al. [25]. The authors consolidated efforts to create a corpus of 4,301 tweets in Saudi dialects called “Tb-SAC.” The corpus was labeled by the authors based on the scale between positive, negative, or neutral tweets, and the article details the data cleaning and pre-processing steps.

Materials and methods

The process of collecting and creating the database for this study went through three major stages, subdivided into several processes, as demonstrated in the flowchart presented in Fig 1. A preliminary analysis of the platform and its operation was carried out, as described below:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Methodological flowchart.

The study’s flowchart.

https://doi.org/10.1371/journal.pone.0316626.g001

A preliminary analysis of the platform and its operation was conducted as described below:

Implementation of the data collection strategy

Strategic use of API resoures.

The data collection process was semi-automated, necessitating the incorporation of new search contexts. Python algorithms were developed to model each primary collection type. During the API assessment and tool implementation phase, it was identified that the “Tweepy” library for Python would streamline the utilization of the Twitter API while maintaining the granularity of desired data fields. Consequently, the data collection system was implemented using these technologies and could be accessed through a terminal.

The system automatically gathered tweets based on specified search contexts (e.g., hashtags from a list, queries from a list, ego networks from a user—tweets made by this user). It was designed to pause when reaching limits and resume automatically upon release. The “token farm” algorithm, elaborated in the following section, was employed to iterate over available API keys, using those not currently at their limits.

Despite being public, the API imposed stringent temporal restrictions on the volume of data collected by a user (10 million tweets per month per token). The study’s intended data collection could not be accommodated under these limitations. Thus, researcher accounts with fewer restrictions were utilized, obtained through a form provided by Twitter. The study was crafted considering the API’s terms of use and limitations applicable to these researcher accounts.

The collection took place through the Twitter API, accessible upon requesting keys (or tokens) with varying access levels. The academic key used in this study is freely available for researchers globally, provided they demonstrate the scientific purpose of using APIs. This key permitted the collection of up to 10 million tweets monthly, with suspension upon reaching this limit until the next month. Additionally, each API request had a maximum limit every 15 minutes, with the request halted until the end of those 15 minutes.

To maximize the database’s potential, the collection involved multiple researcher accounts, totaling six users, each with individual limits. Additionally, accounts with fewer privileges were used for testing purposes and were not involved in the actual data collection.

Another challenge was the contemporaneity of the data and the API’s volume limits being reset multiple times a day. As a result, the data collection was conducted in real-time throughout the period, with the option for retroactive collection if needed.

Token farm.

These limitations were circumvented by utilizing six academic access keys. When one key was suspended, another was activated, optimizing data collection and allowing for the gathering of more tweets per month. This strategy involved rotating keys, ensuring available keys were used while others were in a “rest” period, waiting for a new limit to be released.

Phases of data collection.

With the search contexts defined, the collection was automated. The algorithm ran continuously and required a restart only in the event of external issues.

The algorithm conducted daily collections, fetching content from the day before its execution. Once all tweets for that day were gathered, a JSON stored information indicating nothing else to collect. Subsequently, the algorithm began collecting tweets from previous days that had not yet been included in the collection process. This measure ensured data contemporaneity, allowing the algorithm to seamlessly continue from where it left off. At the end of each day, the collection status, including the information of that day, the applied search context, and the phase of the collection, was stored in JSON. This ensured that the collection would restart from the point it had concluded. Additionally, the context was periodically saved in JSON if the algorithm encountered any complications and needed restarting.

The daily collection consisted of three phases. The first involved a general collection, where all tweets from a specific day were gathered within several requests, each containing a maximum of 500 posts (the API limit). Each request had a specific ID provided by the API, indicating its position in the total collection. The information from these posts, as shown in Table 1, was collected and structured in a DataFrame using the Python library pandas. Information about the authors (Table 2) and media in each post (Table 3) were also collected and stored in separate pandas DataFrames. Consequently, three data files, named “tweets,” “users,” and “media,” were created for posts, users, and media, respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Tweets information.

https://doi.org/10.1371/journal.pone.0316626.t002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Users information.

https://doi.org/10.1371/journal.pone.0316626.t003

The “null” was employed as an indicator of absence in the data files. In the “tweets” file, the fields within the columns referenced_tweet_id, mentions, URLs, hashtags, and media_keys may be null if the post does not fall into the interaction types (retweeted, quoted, or replied_to), lacks mentions, URLs or hashtags in the text body, or contains no media, respectively. In the “users” file, the author_location column may be null if the user has not filled in this editable text field. Additionally, in the “media” file, media_url may be null if the media type is not a photo, as the API does not provide links for videos and GIFs. The field media_view_count may be null if the media type is a photo or GIF, as they do not have a count of views.

The second phase involved searching contexts other than retweets. In this scenario, many collected posts were quotes or replies, necessitating access to the original tweet for context and understanding. This process could be optimized to avoid gathering original tweets already collected. Therefore, all posts in the referenced_tweet_id fields were collected to access the original post and preserve context without. Collection was performed in blocks of 100 IDs, maintaining the same information as in the first phase and stored in the same files. This procedure meant that some older posts, predating the collection day and/or the analysis period, had to be retrieved.

Finally, a third phase focused on the contexts of tweets published in candidates’ accounts. In this phase, all quotes and replies were collected, specifically those linked to the same conversation_id, representing all interactions directly or indirectly related to the initial post of the candidates’ tweets. These collections served as new search contexts. The first phase was applied to each type of interaction collected. The resulting three files were stored in special directories within the folder containing the collection day information. Additionally, a fourth type of file was generated to store the simplified information of the level one ego network generated in this phase, with a file for each interaction (Table 4 describes its structure).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Media information.

https://doi.org/10.1371/journal.pone.0316626.t004

The JSON format used in the first phase was also adopted in the subsequent phases. It served as a central file for controlling which contexts and days had already been collected, managing completion status for each phase, and storing collection status. This facilitated resuming the process from a relatively close point in case of external events, such as machine shutdown or loss of connection, saving time and avoiding token limit issues.

Results

This section presents the main results of a first exploratory corpus analysis. It presents information related to the four main candidates in the 2022 Brazilian presidential elections.

Descriptive measures of the corpus

A total of 282,135,572 posts were collected from Twitter (X), comprising tweets, retweets, quotes, and replies. Because of the functionality of the Twitter API, and a part of the used methods of collection, some of the data includes duplicate objects (due to the API, and intersecting search contexts) or items that date prior to the study’s targeted period (due to the collection of referenced tweets). Among these posts, 176,254,397 were retweets, 65,680,188 were replies, 30,168,951 were tweets, 9,479,217 were quotes, and 552,819 were mixed interactions (posts with both types quoted and replied_to, shown in the graphs as Quoted_replies). Reposts significantly outnumbered other types, highlighting their primary role in content propagation on the social media platform.

Fig 2 depicts multiple peaks throughout the collection period, each corresponding to emblematic events during the study. On August 26, 2022, a significant spike in tweets and retweets occurred following an interview with presidential candidate Lula da Silva on the renowned Brazilian journalistic program, “Jornal Nacional” (Globo TV network). Another peak on August 29, 2022, reflected discussions among Twitter users regarding the first televised debate between presidential candidates the previous night. In October, three notable peaks were observed: the first on October 02, 2022, coinciding with the first round of elections; the second on October 17, 2022, the day after the first debate in the second round; and the third on October 30, 2022, corresponding to the second round of elections. Additionally, two peaks were noted in January: the first on January 01, 2023, marking President Lula da Silva’s inauguration, and the second on January 08, 2023, when supporters of former President Jair Bolsonaro stormed the buildings of the Congress, the Supreme Court, and the Palácio do Planalto, the president’s office.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Number of tweets collected over time.

https://doi.org/10.1371/journal.pone.0316626.g002

Fig 3 illustrates that the search contexts for tweets and retweets related to Lula da Silva and Jair Bolsonaro dominated the corpus, significantly surpassing the number of posts concerning other presidential candidates. Additionally, the search context for Lula’s profile recorded numerous interactions (Table 5), securing the sixth position among contexts with the most tweets.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Number of tweets collected per query and type.

https://doi.org/10.1371/journal.pone.0316626.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 5. Information from interactions.

https://doi.org/10.1371/journal.pone.0316626.t005

A total of 303,683,456 mentions, 75,417,581 URLs, and 26,602,848 hashtags were extracted, providing information about featured accounts, fake news, and prominent hashtags. Additionally, data was collected from 280,404,916 accounts, of which only 3,076,066 were “verified accounts.” Many of these were repeated due to users publishing multiple tweets in a day or across different days and search contexts. Repetitions allowed for the collection of temporal information, aiding in the identification of tweets published in diverse search contexts. The information collected also encompassed 30,673,896 media posts, consisting of 18,502,629 images, 11,505,363 videos, and 665,904 GIFs. Given that many of these media posts were repeated through reposts using the same media as the original post, it is feasible to download still-available media via the provided addresses. This enables the creation of a database of images and videos disseminated during the research period.

Table 6 presents descriptive measures of the corpus, with a notable high standard deviation in candidate-related data. This is a result of search contexts capturing significant spikes in platform activity compared to other days, as depicted in Fig 2. For candidates, this increase is attributed to the surge in tweets close to the first round (October 02, 2022) and second round (October 30, 2022) of the presidential elections. The dates in Table 6 indicate when 25%, 50%, and 75% of the total tweets were collected for each search context.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 6. Descriptive measures of the dataset.

https://doi.org/10.1371/journal.pone.0316626.t006

Analysis of the sample

For the sample analysis, a temporal cut from the database was executed, covering the period from October 27, 2022, to November 02, 2022. This time frame encompasses the pre-election period (October 27, 28, 29, and 30) and the post-election period (October 31, November 01, and 02). Additionally, it represents the period with the highest number of tweets related to the collected hashtags.

Tables 7 and 8 showcases four randomly selected examples of tweet texts from the database within this timeframe. These tweets, diverse in content, consist of emojis, hashtags, and media. They may include news, opinions, and humorous texts, among other types of content. Additionaly, emojis were substituted by standardized literal descriptions for character encoding compatibility reasons. These descriptions are bolded for clarity, and have been translated to portuguese, from it’s standardized english labels, in the portuguese versions of the tweets.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 7. Examples of tweets texts (Portuguese).

https://doi.org/10.1371/journal.pone.0316626.t007

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 8. Examples of tweets texts (English).

https://doi.org/10.1371/journal.pone.0316626.t008

The word cloud, a weighted list model commonly employed for visualizing text data, serves as a vital data visualization tool. Fig 4 illustrates the word cloud generated for the studied time frame by the WordCloud [27] in Python. The cloud is formed based on the frequency of words in the text, and for this result, the NLTK [28] library in Python was utilized to remove irrelevant words (stopwords). The word cloud is presented separately for the pre- and post-election periods and further categorized between candidates Lula da Silva and Jair Bolsonaro.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Wordclouds.

#lula before the election (A), #lula after the election (B), #bolsonaro before the election (C), #bolsonaro after the election (D).

https://doi.org/10.1371/journal.pone.0316626.g004

During the electoral period, words referring to candidate Bolsonaro frequently included terms such as “liar,” “lied,” “despair,” “president,” and “genocidal.” The term “genocidal” was particularly used in reference to Bolsonaro’s anti-vaccine stance during the COVID-19 pandemic. In the post-electoral period, terms reflecting Bolsonaro’s defeat, such as “caiu” (fell), along with nicknames like “bozo” and “frouxonaro,” became prevalent. The nickname “frouxonaro” combines “frouxo” (weak) with “Bolsonaro,” serving as a metaphor for questioning his virility. Other frequently used words included “Nordeste” (Northeast), “ganhou” (won), and “votar” (vote), referring to the winning candidate, Lula.

During the electoral period, the most common words referring to candidate Lula were “president,” “debate,” “government,” and “innocent.” The term “innocent” was emphasized by Lula’s supporters, highlighting the annulment of the charges that led to his arrest. A semantic and interpretative dispute arose over Lula’s arrest and release, with Bolsonaro, his political opponent, using the term “unconvicted.” Words expressing positive feelings, such as “happy,” “good,” and “truth” also appeared frequently. In the post-election period, words like “people,” “won,” “victory,” and “Northeast” were commonly used to celebrate Lula’s victory, especially in the Northeast region of Brazil.

Fig 4 presents Wordclouds separately for the pre- and post-election periods and further categorized between candidates Lula da Silva and Jair Bolsonaro.

Table 9 provides statistical measures for the given time frame. The values indicating the number of different users who tweeted are presented in absolute terms. Considering emojis in texts can be advantageous for tasks such as sentiment analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 9. Table of the database statistical measures.

https://doi.org/10.1371/journal.pone.0316626.t009

Fig 5 depicts a map of Brazil based on data collected solely on October 30, 2022, showcasing the active locations of users who tweeted at least once on that date. Data processing aimed to extract as many locations as possible, given that the profile location, while not mandatory, allows users to input any characters in the text box.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Tweets about Lula and Bolsonaro on October 30, 2022.

The maps were generated using python and the python library “Geopandas” [29], with map data (contours) provided by the same library.

https://doi.org/10.1371/journal.pone.0316626.g005

Discussion

The higher volume of tweets related to candidate Lula during October and November, with over 5 million users making publications about the then-candidate (surpassing just over 4,3 million users posting about Bolsonaro), should not necessarily be interpreted as an indication of greater support on social media platforms for Lula da Silva.

In the 2018 elections, negative propaganda from Bolsonaro’s campaign targeted not only Fernando Haddad, his main opponent from the Workers’ Party (PT), but also former President Lula, who was serving a sentence in Curitiba after being convicted of passive corruption, and the PT itself. Although considered a legitimate resource for providing information and stimulating debate on public policies, candidate proposals, and past achievements [30], negative propaganda took on new dimensions with the emergence of electoral campaigns on digital platforms.

The task of the Brazilian Electoral Justice in monitoring negative propaganda became more complex as multiple actors initiated attacks on the electoral process through social media platforms, adopting a non-official negative campaign strategy.

In this context, Jair Bolsonaro’s campaign and supporter groups not directly affiliated with the campaign adopted the strategy of deconstructing the image of Lula and the PT through social media. In 2022, a consistent effort was made to link candidate Lula with corruption, and the term “thief” was prominently featured in the candidate’s word cloud.

This strategy, at times, relied on controversial facts such as past judicial accusations against the candidate and, at other times, on falsehoods, such as the claim that Lula’s convictions had not been annulled, which the Federal Supreme Court (STF) had done in 2021. Consequently, there was an emphasis on negative propaganda to portray negative attributes of Lula as more concerning than Bolsonaro’s negative attributes, suggesting that issues like corruption and the alleged absence of moral values should be viewed with greater gravity than criticisms of Bolsonaro’s handling of the Covid-19 pandemic.

Based on the exploratory analyses and the quantitative data obtained, the database holds enormous potential for analysis and knowledge acquisition. It serves as a valuable source with unique characteristics stemming from the dynamics of digital social media platforms like Twitter. These social and political data provide a snapshot of the Brazilian political landscape, particularly concerning the 2022 elections. Historically, electoral data with significant informational dimensions have been highly sought after, given their decisive potential in sociopolitical processes and power struggles, carrying profound implications for one or more nations.

The availability of data on the internet with “free” access doesn’t necessarily mean that the “information” is free. Deciding on the analyses, transformations, and visualizations possible with the dataset is not trivial and might not even constitute a countable set of tasks. So far, The work represents only a small fraction of the potential explorations. Hopefully, this effort will pave the way for numerous others to contribute further to the democratization of Brazilian sociopolitical information.

Limitations

Although the goal was to limit the amount of data preprocessing to only what was strictly necessary (leaving further refinement for future work), the possibility of conducting a thorough analysis of the presence of false positives in the final dataset was acknowledged. This would involve identifying the presence of data related to content that does not fit within the sociopolitical scope of the study or does not contribute to the overall understanding of the initially outlined research objectives. However, due to the scale of the dataset produced, there are numerous and varied methodologies that could be employed to address this issue. Most, if not all, include heuristics that may not yield completely accurate results or rely on subjective definitions, choices, and decisions regarding attempts to answer questions such as: ‘Which Tweets are relevant for studies derived from this dataset?’

The potential benefit of applying such preliminary analysis to future studies based on this work is understood, yet for the reasons highlighted, it is concluded that such data preprocessing falls outside the scope of this study, allowing future work to conduct analyses within their specific contexts and leaving the decision-making power of what is relevant in the hands of those who will use the dataset produced. This conclusion is further reinforced when considering that most studies will likely use a subset of the dataset or data pertaining to a specific context (political, for example). In many cases, identifying false positives, given the broad range of potential contexts, may require considering many subtleties. These subtleties and the presence of multiple contexts justify the presence of different methodologies for selection—of which, probably the most appropriate are AI based—and which, to achieve the best results, would require a different input, training data, or fine-tuning, depending on the target context.

Finally, despite this limitation, confidence in the value of this dataset remains, given the scarcity of large datasets like this available in the academic-scientific context; and also in light of the considerations and decisions made by specialists in sociopolitical studies, which led to the preliminary observations and the construction of the queries used for data collection, as outlined in this study. We believe that future users of this dataset will be able to extract that value using existing and well-known methodologies for filtering text-based datasets chosen based on the specific objectives and/or political context that they choose to study.

Data availability

Different strategies for making the data available were considered to ensure secure data storage and accessibility of the research results.

The primary data collected were stored on physical servers at the research group’s Federal University of São Carlos headquarters. This ensures the complete dataset is in a secure location with restricted, controllable, and traceable access.

Regarding the public availability of the data, relevant legal terms were consulted—especially those applicable to the Twitter platform—along with ethical guidelines for research with Big Data and strategies used in scientific studies focused on Twitter with similar legal and availability requirements. It was found that most of the datasets from the 153 articles collected in the systematic literature review were made available on GitHub. All analyzed articles that mentioned ethical issues reported that this type of data collection is based on public data and is exempt from the Research Ethics Committee, which is the case for the present research, following national standards and resolutions. The methods of data availability for the largest corpora described in Table 1 of this article were also analyzed in detail. After thorough analysis and consideration, it was concluded that the most appropriate option would be to adopt a strategy similar to that of Chen et al. [31], who collected 1.2 billion tweets related to the 2020 US elections.

Consequently, the chosen data storage and distribution channel was GitHub [32], commonly used to organize and share software codes and configurations and allows the sharing of text files or compressed files. On GitHub, it is possible to make large volumes of data available, along with descriptions and usage licenses, without the need for payment or fees, as long as the size of each file in the repository does not exceed a certain limit in megabytes (MB).

The dataset produced consists of multiple files (in text format, compressed), organized and classified according to the queries and dates related to the data collection period and the types of objects they represent (users, tweets, or media files). Therefore, to transform the dataset into a publicly available version, the same structure was maintained without any pre-processing, only filtering the content distributed publicly.

To comply with ethical and legal principles, the published data was duly anonymized, removing all fields that could be used to identify users. This procedure prevents and avoids the inference of individual users’ sensitive characteristics by third parties. Therefore, it was decided to make only the IDs of the objects in the dataset available. All information related to each object is linked to its ID and curated by X (formerly Twitter). It can be re-obtained—at the time of writing this article—through the same API used in this study. This process is commonly called “rehydration” of the data [31], and is straightforward once you have the target IDs and access to the API (which is summarized as an access key).

If researchers are interested, additional information related to accounts and users can be requested directly from X through its API in the rehydration process, which will be legally responsible for sharing it. However, this process of obtaining user account data depends on the operation of the X platform and may be subject to prices and fees imposed after its acquisition by Elon Musk. Nonetheless, the data from this research, made publicly and freely available, allows for a variety of analyses and applications. This strategy made it possible to distribute the study’s data while maintaining security, integrity, and compliance with Twitter’s legal terms and the privacy of individuals whose public data may be present in the dataset. Tools that facilitate the rehydration process are available on the GitHub account of the Interfaces research group, along with the dataset. A brief explanation of the process is also available, with links provided in the Data Availability statement of this article. The code developed in Python for this data collection was also made available.

Future research

Following the assessment of the database’s potential for supporting future work, planning its infrastructure, particularly concerning data accessibility and handling, became necessary. While the results of this study already suggest interesting possibilities for use and manipulation, the potential for enhancing these aspects is evident, considering recent computational and technological advances in the context of big data.

Firstly, concerning storage, transitioning to a client-server architecture is believed to ensure better, more standardized data accessibility and increased capacity for parallel data distribution. This change would involve automating access control (currently manual) and adopting proprietary software for query solutions designed for handling large amounts of data in a distributed manner.

In this sense, the software would be installed on a “cluster” type server (computational artifacts managed in parallel) with high processing and storage capacity. This solution is commonplace for processing and distributing big data, as it effectively addresses the complexities associated with working with such data by accommodating various types of uses in the modern information environment.

Therefore, it is expected that, in the future, the database will be accessible through environments employing distributed data storage and processing software, such as Apache Hadoop (Hadoop Distributed File System (HDFS) + MapReduce) [37] and/or Apache Spark (with HDFS or NOSQL database). Alternatively, solutions like Elasticsearch [38], a search and data analysis tool based on the REST and distributed concept, capable of handling a large volume of requests quickly and almost in real time, can be considered.

These software solutions abstract immense complexity, enabling more efficient handling of large amounts of data without the need to implement complex solutions from scratch. This facilitates the configuration of a secure, high-performance data server with extensive options for obtaining data from query languages.

Furthermore, a more appropriate architecture supports the effective retrieval and storage of media data, currently only present as links, due to the significant volume they would occupy if saved locally.

While addressing the storage and accessibility of the database is crucial for the future of this research, the generation of contextualized content holds significance for future studies on social impact topics. Analyzing and studying social and political phenomena and behaviors can contribute to the development of public policies and enhance societal understanding of representativeness, enabling the monitoring of democratic processes. Therefore, it is valuable to extract and create subsets of this database focused on specific contexts related to laws, regulations, mandates, policies, practices, traditions, values, and beliefs at the intersection of social and political life.

For instance, fake news is a critical topic associated with social media, and democracies are increasingly susceptible to its influence. Fake news has been created and distributed to deceive and manipulate, influencing opinion, altering power relations, strengthening hate groups, and fueling prejudice [39].

In this context, there is a need to construct a database specifically for fake news, serving as a benchmark to calibrate, train, and validate specific detection algorithms for Portuguese and for general positive learning algorithms.

Moreover, this new database holds potential for various scientific purposes, including the study of political and behavioral phenomena like echo chambers, accentuating political polarization, and homophily, defined as the tendency of individuals to establish relationships based on shared interests. Such phenomena are closely linked to the spread of misinformation [40].

It is worth noting that there are numerous possibilities for creating contextualized and annotated subsets, such as selecting tweets related to hate speech and toxic content, especially when associated with cyberbullying and other important social impact topics.

Finally, as tweets can be associated with images and videos, we aim to make media related to our database available. This information can prove valuable, particularly in the field of computer vision, which focuses on analyzing, interpreting, and extracting relevant information from images and videos, aiming for accurate and efficient descriptions.

Conclusion

While Twitter is generally a “public” access platform and fits into big data standards, extracting valuable information is not trivial due to the volume, speed, and heterogeneity of data. This study concludes that acquiring informational value requires expertise not only in sociopolitical areas but also in computational and informational studies, highlighting the interdisciplinary nature of such research.

As long as we cannot democratically access the dynamics of certain important processes (in terms of democratic informational access), society will not be able to enjoy the complete democratic process.

Consequently, well-documented and scientifically studied datasets are crucial from a democratic perspective, offering society comprehensive information that illuminates historical political events.

Acknowledgments

National Council for Scientific and Technological Development (CNPq) under grant number 420025/2023-5.