Frequently asked questions

Context annotations

The questions below are specific to the context annotations element of Tweet annotations. For more details, please see the Overview page.

How does Twitter context annotations work?

Twitter classifies Tweets semantically, meaning that we curate lists of keywords, hashtags, and @handles that are relevant to a given topic. If a Tweet contains the text we’ve specified, it will be labeled appropriately. This differs from a machine learning approach where a model is trained specifically to classify text (in this case, Tweets) and produce a probability score alongside the output/classification.

How do I know that your data is complete and trustworthy?

Twitter's annotations are curated by domain experts using research and QA processes that have been refined over the course of several years. The process is supported by custom tooling to scale data tracking as far as we are able to maintain excellent precision and recall. In addition, our data is audited regularly by an internal team, having received a precision score of ~80% for the past several quarters.

 How do you ensure precision?

Team members QA our entities on a daily basis to ensure high precision and recall. Additionally, our work is audited quarterly by an internal team, which manually reviews 10,000 Tweets across all of our domains to calculate a precision score.

How do you decide what to track?

For some domains, like sports and TV, we rely on automated ingest to build out our graph. In the News domain, we track data around stories published by the Twitter Moments team. Otherwise, the team uses a variety of research methods to identify topics to track that garner a high amount of conversation on the platform.

What historical support is available with Tweet Annotations?

Data tracking begins the moment an entity is published; therefore, we do not annotate Tweets that were published prior to a given entity being tracked. For example, if an upstart brand/company is added to the taxonomy, we will not retroactively annotate Tweets about that brand prior to when the annotation was added.

Is Twitter able to annotate Tweets in non-english languages? If so, which languages and does the coverage of Tweets being annotated change?

Yes. Language coverage can vary depending on the domain and the market. English and Japanese are included in the majority of the biggest entities. Below, is a list of the languages and main markets that are covered today:

1. English (US, UK)
2. Japanese (Japan)
3. Portuguese (Brazil)
4. Spanish (Argentina, Mexico, Spain)
5. Hindi (India)
6. Arabic (Saudi Arabia)
7. Turkish (Turkey)
8. Indonesian (Indonesia)
9. Russian (Russia)
10. French (France)

Coming soon (~H2 2021):

1. German (Germany)
2. Tamil (India)

Below is a table of the top 15 countries ordered by the most coverage of annotated Tweets:

What underlying "semantics" does Twitter rely on to annotate a Tweet?

Tweet annotations consist of the following semantics to annotate a Tweet:

• Accounts - we can annotate tweets from a given handle or mentioning this handle
• Hashtags
• Keywords/phrases

For customers that are familiar with the filtered steaming APIs such as PowerTrack, the semantics used by annotations are similar in principle to the boolean rules defined to filter a stream of Tweets. If a Tweet matches the underlying semantic conditions, it will be tagged accordingly.

Why do some Tweets have entities associated with them while others do not?

The goal is to annotate as many Tweets as possible; however, there are several reasons why some Tweets are not annotated:

• Some Tweets aren't semantically rich enough to be labelled and can't be tagged with our current annotation rules
• Some Tweets aren't topical
• The Tweet is about a very ephemeral topic that's not in our graph
• We don't cover the language/market
• We cover the language/market but we're missing a topic or a specific term/account/hashtag related to a topic we already track"

When there are multiple domains (for example, [3,30]), does the Entity ID remain the same?

An entity can be part of multiple domains. The domain IDs will change but the entity ID remains the same. Donald Glover is a person (domain 10), an actor (domain 56)  and a musician (domain 54) but his entity ID is still 875072662527029248.

Do you have an established timeline for show/movie tracking? In other words, how long is a show/movie tracked before/after release?

Tracking starts a month prior to the release. For popular blockbusters, like a Marvel movie, we can start tracking them as soon as they start teasing about an upcoming release.

Do movies have a locale filter similar to the one for TV shows?

No, they do not.