Facebook's AI tech can spot hate speech by reading text from photos and videos

'Rosetta' system uses machine learning to understand text in images and videos so the social network can monitor hateful content

penguin meme - text being recognised

Facebook has created an artificial intelligence-powered system for reading text from photos, that it will use to spot hateful and inappropriate memes.

The system is called 'Rosetta' and uses machine learning to identify text in images and videos that can be then transcribed into something that is machine readable.

The social network has described the system in great detail in a blog post and say it is being put to use to "automatically identify content that violates our hate-speech policy".

"Understanding the text that appears on images is important for improving experiences, such as a more relevant photo search or the incorporation of text into screen readers that make Facebook more accessible for the visually impaired," Facebook said.

"Understanding text in images along with the context in which it appears also helps our systems proactively identify inappropriate or harmful content and keep our community safe."

Facebook has over two billion users and monitoring all the content posted daily is virtually impossible. However, a significant number of its users (and social media users in general) post content with written text overlaid, such as a meme, and the social network has built and deployed what it calls a "large-scale machine learning system".

The social network giant says the system can extract text from more than a billion public Facebook and Instagram images and video frames, in a number of languages, and then use a text recognition model to understand the context of the text and image together.

Rosetta has been trained on a vast data set of what Facebook calls "human-annotated" public images with text and their locations and also text on wider public images such as posters and menus.

Facebook says that Rosetta has already been widely adopted within various teams at the company and with Instagram, and noted that the text extracted from images is also being used in various machine learning models, such as those to improve its photo searching features.

Featured Resources

Become a digital service provider

How to transform your business from network core to edge

Download now

Optimal business results with the cloud

Evaluating the best approaches to hybrid cloud adoption

Download now

Virtualisation that enables choices, not compromises

Harness the virtualisation technology that's right for your hybrid infrastructure

Download now

Email security threat report 2020

Four key trends from spear fishing to credentials theft

Download now

Recommended

Parler suffers data leak before being taken offline
social media

Parler suffers data leak before being taken offline

12 Jan 2021
Trump's TikTok ban hits another roadblock
social media

Trump's TikTok ban hits another roadblock

9 Dec 2020
People can’t resist opening LinkedIn phishing emails
Security

People can’t resist opening LinkedIn phishing emails

16 Nov 2020
Snapchat now allows you to display subscriber numbers
social media

Snapchat now allows you to display subscriber numbers

4 Nov 2020

Most Popular

150,000 arrest records accidentally deleted from police database
data management

150,000 arrest records accidentally deleted from police database

15 Jan 2021
How to recover deleted emails in Gmail
email delivery

How to recover deleted emails in Gmail

6 Jan 2021
What is a 502 bad gateway and how do you fix it?
web hosting

What is a 502 bad gateway and how do you fix it?

12 Jan 2021