TBCOV: Two Billion Multilingual COVID-19 Tweets with Sentiment, Entity, Geo, and Gender Labels

Welcome to the code repository for the TBCOV dataset. This code repository offers several scripts helpful to hydrate and process the shared dataset.

The TBCOV dataset comprises more than two billion tweets from 218 countries worldwide. The following map shows worldwide tweets normalized by the total population from each country per 100,000 persons.

Data descriptor for base release files

Attribute	Type	Description
tweet_id	Int64	The integer representation of the unique identifier a tweet. This number is greater than 53 bits and some programming languages may have difficulty/silent defects in interpreting it.
date_time	String	UTC time when the tweet was created.
lang	String	ISO-6391 Alpha-2 language code consisting of two characters.
user_id	String	Represents the id of the author of the tweet.
retweeted_id	Int64	If the tweet is a retweet, the retweeted_id represents the id of the parent tweet.
quoted_id	Int64	If the tweet is a quoted tweet, the quoted_id represents the id of the parent tweet.
in_reply_to_id	Int64	If the tweet is a reply to an existing tweet, the in_reply_to_id represents the id of the parent/original tweet.
sentiment_label	Int64	Represents the sentiment label values: -1 (negative), 0 (neutral), 1 (positive).
sentiment_conf	Float	Represents the confidence score of the sentiment classifier for a given sentiment label to a tweet.
user_type	String	The user types represents the identified type of the user such as person, organization, location, etc.
gender_label	String	One character code representing the identified gender of the users. F represents "female" and M represents "male" user types.
tweet_text_named_entities	Dictionary array	Named-entities (persons, organizations, locations, etc.) extracted from tweet text are provided in this attribute in the array of dictionary format.
geo_coordinates_lat_lon	Float	GPS coordinates in the latitude, longitude format retrieved from the user's GPS-enabled device.
geo_country_code	String	Two characters country code learned through resolving the GPS coordinates (latitude, longitude).
geo_state	String	The name of the state/province learned through resolving the GPS coordinates (latitude, longitude).
geo_county	String	The name of the county learned through resolving the GPS coordinates (latitude, longitude).
geo_city	String	The name of the city learned through resolving the GPS coordinates (latitude, longitude).
place_bounding_box	Float	Twitter provided bounding boxes representing place tags.
place_country_code	String	Two characters country code learned through resolving the place bounding boxes.
place_state	String	The name of the state/province learned through resolving the place bounding boxes.
place_county	String	The name of the county learned through resolving the place bounding boxes.
place_city	String	The name of the city learned through resolving the place bounding boxes.
user_loc_toponyms	Dictionary array	Toponyms recognized and extracted from the user location field provided as an array of dictionary.
user_loc_country_code	String	Two characters country code learned through resolving the user location toponyms.
user_loc_state	String	The name of the state/province learned through resolving the user location toponyms.
user_loc_county	String	The name of the county learned through resolving the user location toponyms.
user_loc_city	String	The name of the city learned through resolving the user location toponyms.
user_profile_description_toponyms	Dictionary array	Toponyms recognized and extracted from the user profile description field provided as an array of dictionary.
user_profile_description_country_code	String	Two characters country code learned through resolving the recognized user profile description toponyms.
user_profile_description_state	String	The name of the state/province learned through resolving the recognized user profile description toponyms.
user_profile_description_county	String	The name of the county learned through resolving the recognized user profile description toponyms.
user_profile_description_city	String	The name of the city learned through resolving the recognized user profile description toponyms.
tweet_text_toponyms	Dictionary array	Toponyms recognized and extracted from the tweet full_text field in the dictionary array format.
tweet_text_country_code	String	Two characters country code learned through resolving the recognized tweet text toponyms.
tweet_text_state	String	The name of the state/province learned through resolving the recognized tweet text toponyms.
tweet_text_county	String	The name of the county learned through resolving the recognized tweet text toponyms.
tweet_text_city	String	The name of the city learned through resolving the recognized tweet text toponyms.

Tweets hydration

The tweets hydration process fetches full tweet content from Twitter using tweet-ids. To assist users with hydrating TBCOV tweets, this code reposity provides a tool written in the Java language that takes tweet-ids as input and retrieves full tweet content from Twitter APIs. More details and a usage guide of the Tweets hydrator are available here.

Preprocessing

Different types of preprocessing were applied on different attributes before using them for any analysis. The preprocessing is important to replicate results. The code reposity provides several scripts used to preprocess different fileds. The preprocessing scripts are avaialablel here.

Meta-data file

The meta-data file provides a convenient and faster way to retrieve tweets from the base files. The meta-data file holds the start and the end tweet-id of all base files. So, given a tweet-id file (e.g., a language or a country), the provided script determines which base files to parse to retrieve tweets matching the ids instead of parsing all two billion tweets.

meta_file_monthly_ids_range.tsv file lists range of tweet IDs (between Start_id and End_id) contained in the specific monthly base file as follows:

File_name	Start_id	End_id
february_2020_f1.tsv	1223395535882768385	1231201257739649025

Parsing using meta-file

In the folder parsers, there are two scripts that are needed to extract the tweet details from base release files, given a specific language or country IDs file.

• meta_file_parser.py requires two arguments as input.

1. Country/Language IDs file
2. meta_file_monthly_ids_range.tsv

Sample to run the script is as follows:

python meta_file_parser.py [IDs file] meta_data/meta_file_monthly_ids_range.tsv

And it creates an output file name required_monthly_files.txt The contents of the above file look something like this:

february_2020_f3.tsv
february_2020_f2.tsv
february_2020_f1.tsv

Download the above files from Crisis-NLP TBCov It lists the monthly base files required to extract the full data from IDs files. Make sure to download the full monthly base zip file to get the individual files stated in the above required_monthly_files.txt file.

Each line indicates which monthly base file is required for download so that it can be used to extract tweet details with the help of the next script.

• base_file_data_extractor.py requires three arguments as input.

1. required_monthly_files.txt which the output of the previous script
2. Country/Language IDs file
3. Base release files path (expects '/' at the end - for example - /home/downloads/)

Sample to run the script is as follows: python base_file_data_extractor.py required_monthly_files.txt test_for_meta_parsing.txt '/some/path/'

The output will be a .tsv file which will have the same formate as montly base files.