📄 Large Arabic Twitter Dataset on COVID-19

Summary

Arabic has roughly 300 million native speakers and a very active Twitter presence, yet the COVID-19 tweet collections released early in the pandemic contained little Arabic material. The authors fill that gap with what they present as the first Twitter corpus devoted solely to Arabic posts about the pandemic. Starting on 1 January 2020, they streamed tweets through the Twitter streaming API using the Tweepy library. A tweet was captured when its text matched a curated list of common Arabic COVID-related keywords or one of several tracked hashtags. The list grew over time as new terms appeared, and the date each term was added is recorded. They kept the full tweet objects and removed tweets judged unrelated to the pandemic. By 15 April 2020 the collection held about 3.9 million tweets. Twitter's redistribution rules forbid sharing full content, so the public GitHub release contains only tweet IDs, which users can rehydrate. The resource is meant to support research on public behaviour, information sharing, misinformation and rumour spread in the Arabic-speaking world.

Key points

• The collection ran from 1 January to 15 April 2020 via the Twitter streaming API and Tweepy, giving 3,934,610 tweets and an average of 77,471 per day.
• There are 24 Arabic keywords (Table 1), each with its own tracking start date. Early terms include Corona, Wuhan and mask. Later additions include curfew (15 March) and ventilator, cough and "one and a half meters" (1 April). Sixteen hashtags were also tracked (Table 2).
• Table 3 reports 219 geotagged tweets, 3,934,235 original tweets and 375 retweets. The authors also say retweets rose sharply in late March, and these two statements are hard to reconcile.
• Only tweet IDs are published, at github.com/SarahAlqurashi/COVID-19-Arabic-Tweets-Dataset, to comply with Twitter policy. Full objects can be retrieved with tools such as Hydrator.
• The authors intend to continue updating the dataset and to examine how various groups react to the pandemic.

Interesting ideas

• I1. The query vocabulary is logged with a start date for every keyword, so later users can see which terms could have produced a tweet at a given time and avoid mistaking vocabulary expansion for real trends. — For Mova (learn): Silver-data and teacher corpora could store the date and version at which each source, filter or rule entered the pipeline, so evaluation shifts can be traced to coverage changes rather than model changes.
• I2. Search terms are allowed to evolve as the event unfolds, moving from "Wuhan" and "mask" to "curfew" and "social distancing". Fixed keyword lists can miss emerging language, and this design avoids that. — For Mova (world): Compiled knowledge such as ontology entries and value lexicons could carry time-of-validity tags, so the world model knows that a term or entity relation only became salient after a certain point.
• I3. The authors publish identifiers instead of content and let users rehydrate the data, which keeps a large resource shareable despite licensing limits. — For Mova (learn): Evaluation sets built from restrictively licensed text could be distributed as pointers plus Mova's own derived annotations (UD trees, frames), keeping the explainable layers shareable without redistributing source text.
• I4. Simple summary counts, such as original versus retweets, geotagged share and daily averages, work as sanity checks. Here they expose an internal inconsistency with the narrative, showing why such tables should be cross-checked against claims. — For Mova (learn): Automated consistency checks between reported corpus statistics and pipeline logs could serve as a negative control, flagging silent data bugs before training on silver data.