The repository contains an ongoing collection of tweets IDs associated with the novel coronavirus COVID-19. The dataset contains Tweets’ ids dating back to January 22th, 2020. The Twitter’s search API was used to gather historical Tweets from multiple continents in multiple languages that contained a given keyword (i.e., coronavirus, virus, covid, ncov19, ncov2019). In order to comply with Twitter’s Terms of Service, only the Tweet IDs of the Tweets gathered are released for non-commercial research use only.
The dataset is also on Kaggle:Covid19-tweets-dataset
Citation
Christian Lopez, Malolan Vasu, and Caleb Gallemore (2020) Understanding the perception of COVID-19 policies by mining a multilanguage Twitter dataset. arXiv:cs.SI/2003.10359,2020 https://arxiv.org/abs/2003.10359
The tweet-IDs are organized by keywords as follows:
-
Tweet-ID files are stored in folders that indicate the keyword used (i.e., coronavirus, virus, covid, ncov19, ncov2019).
-
The Tweet-ID file is a
.txtcontaining individual Tweet-IDs stored as a comma-separated-list. The file names follow the naming convention of :[Keyword]_yyyy_mm_dd.txt(e.g.,coronavirus_2020_01_22.txtcontains the ids of tweets mentioning the keyword ‘coronavirus’ tweeted on January 22th, 2020). -
The Tweets_ID_by_language folder contains a series of
.txtfiles with the tweets ID and their language organized by keyword (see Twitter language codes). The file names follow the naming convention of :[mm_dd_yyyy]_[Keyword].txt(e.g.,coronavirus>Apr_03_2020_coronavirus, contains tweets id and their language from all the tweets collected on April 3rd, 2020 using the keyword coronavirus). -
Since the Twitter API returns tweets in UTC, all tweet-ID folder and file names are all in UTC as well.
- Not all keywords were tracked from the very beginning. As news about the novel coronavirus spread, additional keywords were added to the search list. The keywords used for search tweets are: virus and coronavirus since 22 January, ncov19 and ncov2019 since 26 February, and covid since 7 March 2020.
- Since the Twitter API can provide tweets up to 7-days in the pass, there is a lag of 7-days in the dataset to ensure as much tweets as possible are sampled.
- Our dataset does not capture every single tweet on Twitter relating to the keywords as our data retrieval is restricted by the Twitter API at 45,000 tweets every 15 minutes.
- We currently have some descriptive statistics about our dataset displayed in the paper associated with it. We are working on automatically updating these statistics with every update of the dataset.
- Consider using tools such as the Hydrator and Twarc to rehydrate the tweet-IDs i.e. to fetch tweet data associated with the tweet-IDs using Twitter’s API.
- Tweets deleted after our initial collection may still be present in our dataset, but their details can no longer be retrieved by using Twitter’s API.
- Chen et al., 2020 has published a dataset, similar to ours, containing tweet-IDs relating to the novel coronavirus. For those looking for an even larger dataset than ours, merging the two datasets is a valid option.
- Only tweets in English were collected from 22 January to 31 January 2020, after this time the algorithm collected tweets in all languages.
- You can notice that there are fewer tweets relating to the novel coronavirus present in our dataset in the first few weeks than in the following weeks. This can explained by the fact that our data collection involved tracking other keywords unrelated to the coronavirus during the first few weeks (other projects). This meant that the number of coronavirus-related tweets were limited at a fraction of the total API limit. Once unrelated keywords were removed from our search options, the limit of coronavirus-related tweets became the total API limit, hence resulting in more tweets in our dataset.
- Our dataset is currently missing tweed-ID entries for Feb 06 and 07, 2020. This data was lost due to technical errors. However, tweet data was retained and hence our statistics were not affected by this loss of tweet-IDs. We currently are working on retrieving the lost tweet-ID data.
- Our dataset is currently missing tweed-ID entries for March 24th keywords: virus, coronavirus, covid, nvoc2019 (i.e., only ncov19 are present in the data set). This data was lost due to technical errors.
The notebook Automatically_Hydrate_TweetsIDs_COVID190.ipynb will allow you to automatically hydrate the tweets-ID from our COVID19_Tweets_dataset GitHub repository.
You can run this notebook directly on the cloud using Google Colab (see how to tutorials) and Google Drive.
In order to hydrate the tweet-IDs using TWARC you need to create a Twitter Developer Account.
The Twitter API’s rate limits pose an issue to fetch data from tweed-IDs. So, we recommended using Hydrator to convert the list of tweed-IDs, into a CSV file containing all data and meta-data relating to the tweets. Hydrator also manages Twitter API Rate Limits for you.
For those who prefer a command-line interface over a GUI, we recommend using Twarc.
Using Hydrator
Follow the instructions on the Hydrator github repository.
Using Twarc (CLI)
Follow the instructions on the Twarc github repository.
-
First follow instructions for installation.
-
Then, obtain your Twitter API token (apply for a Twitter developer account).
-
Configure twarc with your API token by following the instructions for configuration.
-
Follow instructions in the Hydrate section. Hydrated tweets are stored in jsonl files.
The following data is from Tweets colectled until Apr 18, 2020. (v1.5)
Number of Tweets: 65388324
Number of Original Tweets: 14597888
Number of Re-Tweets: 50790436
Tweets Breakdown by Month:
| Month | Number.of.Tweets |
|---|---|
| Jan | 724877 |
| Feb | 2994768 |
| Mar | 27414279 |
| Apr | 34254400 |
Tweets Breakdown by Search/Keyword:
| Keyword | Number.of.Tweets | Percentage |
|---|---|---|
| coronavirus | 22823353 | 34.90 |
| virus | 22059400 | 33.74 |
| covid | 20342766 | 31.11 |
| ncov2019 | 82870 | 0.13 |
| ncov19 | 79935 | 0.12 |
Tweets By Language:
| Language | Number.of.Tweets | Percentage |
|---|---|---|
| English | 38455759 | 59.41 |
| Spanish; Castilian | 11203759 | 17.31 |
| Portuguese | 4309676 | 6.66 |
| Bahasa | 2684019 | 4.15 |
| French | 2514900 | 3.89 |
| Italian | 880907 | 1.36 |
| Thai | 829668 | 1.28 |
| Japanese | 663834 | 1.03 |
| Hindi | 535609 | 0.83 |
| Turkish | 483823 | 0.75 |
| Catalan; Valencian | 401929 | 0.62 |
| Tagalog | 383925 | 0.59 |
| German | 371524 | 0.57 |
| Dutch; Flemish | 182799 | 0.28 |
| Other | 831355 | 1.28 |
Number of tweets a day by language for the keyword ‘coronavirus’.
Geolocated tweets for the keyword ‘coronavirus’
Average sentiment of tweets for the observation period grouped by keyword.
For questions about the dataset, please contact Dr. Christian Lopez at lopezbec@lafayette.edu, Dr. Caleb Gallemore at gallemoc@lafayette.edu, or Malolan Vasu at vasum@lafayette.edu.
This dataset is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International Public License (CC BY-NC-SA 4.0). By using this dataset, you agree to abide by the stipulations in the license, remain in compliance with Twitter’s Terms of Service, and cite the following manuscript:
Emily Chen, Kristina Lerman, and Emilio Ferrara. 2020. #COVID-19: The First Public Coronavirus Twitter Dataset. arXiv:cs.SI/2003.07372, 2020