Machine learning, neural networking and deep learning are terms you generally hear thrown around in the top echelons of the artificial intelligence world, so hearing them from Facebook again is quite fitting. In particular, they’ve created a brand new artificial intelligence, based on deep learning, named, aptly, DeepText. The newborn A.I. is being hailed as a “text understanding engine”, able to check out text and figure out the context, content, and even classification, all based on deep learning algorithms that let it rack up experience as it examines more and more posts over time. The kicker here is that Facebook is only just announcing this new A.I. and says that it’s already a stone’s throw from being at human level in understanding text, which is an extreme accomplishment, even by tech giant standards.
According to Facebook, this A.I. will be able to conceptualize, understand and categorize thousands of posts per second across the service, armed with support for 20 languages. The bot is actually capable of analyzing text on a language-agnostic level through essentially “learning” the very core of human language – the patterns, variations and subtleties that permeate all speech. Even with only 20 languages under its belt, using this method of deep neural learning to subvert the need for traditional “teaching” of each language to the robot is an incredible feat. The A.I. is also able to detect words that may mean the same thing based on how close they are in spelling, context and use, such as “bro” and “brother”. By this same logic, the bot can understand those who may have a hard time getting their spelling right – new or misspelled words are analyzed deeply in context, rather than matched as closely as possible with a word that the robot already knows the spelling and definition of.
DeepText is made to be able to understand slang and disambiguation, allowing it to learn more and more about those two subjects unaided, unlike traditional A.I. that have to be fed new terms, thanks to the magic of neural networks and deep learning. To put a very long concept as short as possible, the A.I. is able to check through assigned content, make inferences based on context, and match that up to examples like sentences and concepts that it already knows. This means that it could encounter somebody posting about Marge krumping, and have no idea what’s going on at first, but then look at data on krumping that it has from people posting about dancing, data on various Marges in its database, and when it runs across the given meme from The Simpsons, everything clicks into place. DeepText is already being utilized in tests in some Facebook services, but plans to improve its ability to join image and text contextual recognition, understand people’s interests and integrate new neural networking designs are all in motion.