Facebook has developed a new AI called DeepText that it claims can understand what people are writing about with "near-human accuracy." Oh, and it can do that several thousand times per second in over 20 languages.
DeepText was announced in a blog post today, with Facebook engineers explaining how it works and how it's rolling out. The new AI is ultimately meant to let Facebook understand any writing that it comes across. It intends to use that ability to make better recommendations to people, both about related posts and related services available on Facebook; on top of that, Facebook could also use DeepText to identify spam or harassment and filter those comments out immediately.
So far, DeepText is mostly good for hailing cabs
It hasn't gotten quite that far, of course. For now, DeepText can only understand the context of a post — that is, the subject it's about or the general intent. So far, it's being put to use in two ways: in Facebook Messenger, it's supposed to understand when someone says they need a ride, allowing Facebook to offer to hail them a cab. DeepText is also being used to identify when people are trying to sell things on Facebook, allowing Facebook to recommend the use of its merchant tools. It's a pretty basic start, but it's easy to imagine how this can expand as DeepText learns to identify other requests.
As its name suggests, DeepText has been able to get this far using deep learning. Facebook says this allows DeepText to have a better grasp on meaning than a traditional AI, which wouldn't be as good at linking together topics. Going forward, Facebook hopes to use DeepText to better understand the individual people, places, and events referred to within posts and to use that to offer even more personalized recommendations.