We have been actively collecting tweets since January 28, 2020, leveraging Twitter's streaming API [20] and Tweepy [21] to follow specific keywords and accounts that were trending at the time. When we started collecting tweets, we also used Twitter's search API [22] on the same keywords to gather related historical tweets. Thus, the earliest tweets in our collection date back to January 21, 2020. Since then, we have incrementally added keywords and accounts to follow based on the conversations occurring on Twitter at any time. We have collected over 72 million tweets from inception to March 21, 2020, constituting roughly 600 GB of raw data, and are still collecting data to this day.

Our collection relies upon publicly available data and is hence registered as IRB (institutional review board) exempt by the University of Southern California IRB (approved protocol UP-17-00610). We release the data set with the stipulation that those who use it must comply with Twitter’s Terms and Conditions [23].

By continuously monitoring Twitter's trending topics, keywords, and sources associated with COVID-19, we did our best to capture conversations related to the outbreak.

Twitter's streaming API returns any tweet containing the keyword(s) in the text of the tweet, as well as in its metadata; therefore, it is not always necessary to have each permutation of a specific keyword in the tracking list. For example, the keyword “Covid” will return tweets that contain both “Covid19” and “Covid-19.” We list a subset of the keywords and accounts that we are following in Tables 1 and 2, respectively, along with the date we began tracking them. There are some keywords that overlap due to an included keyword being a substring of another, but we included both for good measure. The keyword choices in the current data set are all in English, so there is a heavy bias toward English tweets and events related to English-speaking countries. Due to the evolving nature of the pandemic and online conversations, these tables will expand as we continue to monitor Twitter for additional keywords and accounts to add to our tracking list.

Our data collection will continue uninterrupted for the foreseeable future. As the pandemic continues to run its course, we anticipate that the amount of data will grow significantly. The data set is available on GitHub [24] and is released in compliance with the Twitter's Terms and Conditions, under which we are unable to publicly release the text of the collected tweets. We are, therefore, releasing the Tweet IDs, which are unique identifiers tied to specific tweets. The Tweet IDs can be used by researchers to query Twitter’s API and obtain the complete tweet object, including tweet content (text, URLs, hashtags, etc) and authors’ metadata. This process to retrieve the full tweet object from Twitter starting from a Tweet ID is referred to as hydration. There are several easy-to-use tools that have been developed for such purposes, including the Hydrator [25] and Twarc [26], but one could also directly use Twitter’s API to retrieve the desired data. This data set can also be found on Harvard Dataverse [27]. Table 3 displays basic statistics, including collection period and number of tweets in that respective release, for all current releases (as of May 15, 2020).

There are a few known gaps in the data, which are listed in Table 4. Due to Twitter API restrictions on free data access, we were unable to recover data from the listed times, as Twitter only provides free access to tweets returned from their streaming API from the past week. To request access, interested researchers will need to agree upon the terms of usage dictated by the chosen license.

All of the Tweet ID files are stored in folders that indicate the year and month the tweet was posted (YEAR-MONTH). The individual Tweet ID files each contain a collection of Tweet IDs, with the file names all beginning with the prefix “coronavirus-tweet-id-” followed by the year, month, date, and hour the tweet was posted (YEAR-MONTH-DATE-HOUR).

We note that if a tweet has been removed from the platform, researchers will not be able to obtain the original Tweet.

We present an initial analysis of our collected data set that verifies that Twitter discourse statistics reflect major events at the time, and leverage Business Insider [29], NBC [30], and CNN [31] released timelines to identify these events of interest during the development of the COVID-19 pandemic. In some of these analyses, there is a dip on March 2, 2020—this was due to internet connectivity failures throughout that specific day. Our discussion is based on analysis done on tweets from release v1.2 (January 21, 2020 to March 31, 2020), while the most recent release is v1.8.

We tracked the frequency of COVID-19-related hashtags, specifically those that contain the substrings “wuhan,” “coronavirus,” and “covid” throughout our collection period (Figure 1). We can see that while hashtags with the substring “coronavirus” consistently remain a more heavily used hashtag in our data set, the hashtag usage spiked on the day the WHO declared COVID-19 a global public health emergency; it also spiked on the day the United States announced the first COVID-19-related death [2]. We also did not see hashtags referencing “covid” being used until February 11, 2020, when the WHO announced “COVID-19” as the official name for the novel coronavirus disease. The keyword “wuhan” in hashtags experienced consistent usage until late February, then steadily declined, which reflects the decrease in cases in China and the global spread of the virus.

We then examined the percentage of total tweets posted in different languages (Figure 2). Although English is the most prominent language in our data set, we excluded English from this analysis to better visualize tweet activity in countries that experienced COVID-19 outbreaks earlier in the timeline. In particular, we found that Japanese tweet activity increased steadily after the cruise ship Diamond Princess was quarantined off the coast of Yokohama, Japan, with a peak around the time when passengers began to disembark [32].

There was also a significant spike in tweets from Italy when the first case related to COVID-19 was reported in Lodi, Italy, and first death was seen in Veneto [33]. We also observed a peak in the percentage of Spanish tweets after the first COVID-19 case in Spain was announced on February 1, 2020 [34] and a steady increase in the percentage of Spanish tweets after reports of the first COVID-19-related death began to emerge (the death itself occurred on February 13th, but the cause was diagnosed postmortem) [35].

Verified users on Twitter have been identified by Twitter as accounts of public interest and are verified to be authentic accounts [36]. We observed that the verified accounts, which include news sources and political figures, are the most active when major events occur, as seen in Figure 3. This is to be expected since influential figures and news sources often weigh in and report on breaking news in real time using Twitter as a platform to amplify their messaging. As the United States also drives much of the discourse on Twitter, it is therefore unsurprising that there is a major spike in activity from verified users when the country experienced its first COVID-19-related death.

There are several limitations to our data set. We collect our data set leveraging Twitter’s free streaming API, which only returns 1% of the total Twitter volume, and the volume of tweets we collected continues to be dependent on our filter endpoint and network connection [37].

While our data set is a multilingual data set, containing tweets in over 67 languages, the keywords and accounts we have been tracking and continue to track have been mostly English keywords and accounts. Thus, there is a significant bias in favor of English tweets in our data set over tweets in other languages.

Despite these limitations, our data collection gathers over 1 million tweets a day from the 1% of tweets available to us through Twitter’s API, and our data set contains on average 35% non-English tweets. Our collection begins in late January, capturing tweets during many major developments, and we plan on continuing collecting tweets for the foreseeable future.