Hate speech is one of the worst forms of unwanted content on our products, and it’s clear the solutions will require a combination of deep human expertise and powerful technology. Facebook’s Community Standards Enforcement Report, released today, shows the advances we’ve made on both sides of this challenge, and highlights progress that’s still to come.
You can see some of those advances here — there’s been a rapid evolution. Starting in 2015, our first early computer vision classifiers were able to find one type of violation in images only. Over time we’ve expanded to multiple types of content: images, text, video, and text in images via OCR. We’ve also expanded our systems to multiple languages, and multiple violation types. These new systems were powered by advances in self-supervised learning, multi-modal learning, and multi-lingual learning coming out of our AI research labs.