Facebook Inc. (FB) announced in a blog post an artificial intelligence system called DeepText that can recognize with “near-human accuracy” users’ needs based on their posts on the site. The idea is to help them reach the specific audience they intend to communicate with through a certain publication, as well as to show them the content they are actually interested in.
The company founded by Mark Zuckerberg said it is still working on the new technology and the applications it has the potential to cover. The deep learning system is based on original ideas by Ronan Collobert and Yann LeCun from Facebook AI Research.
The project’s leaders are the engineers Ahmad Abdulkader, Joy Zhang and Aparna Lakshmiratan. They want to filter out spam and other undesirable content in order to make the user experience more enjoyable so people spend more time on the site sharing the content they love.
While Traditional Natural Language Processing (NPL) helps a computer algorithm read text by converting words into numbers, deep learning uses “word embeddings” to keep the exact meaning of the texts intact, given the complexity of human language.
DeepText can read “I just got out of the taxi” and “I need a ride” and understand the different meanings to know when to help users request an Uber, for example. Another possible application could be the detection of a user’s intention to sell any item and consequently format post with all the product details so it looks different from the rest of the information on the News Feed and catch the attention of most contacts.
The social networking site’s newest AI system has been designed to analyze several thousand posts per second in 20 languages. About 400,000 new stories and 125,000 comments on public videos, text posts and images are shared each minute on the site, and DeepText could recommend users the content they care about among such a large pool of information.
Facebook explained that it is necessary to teach the computer to understand word-sense disambiguation and slang in order to make it nearly human in terms of text comprehension. For example, if someone writes “I love apple” the system must be able to determine whether the text is about the fruit or the brand.
DeepText could also be used to make abusive posts invisible and prevent the consequences of harassment caused to many users of Facebook. If it can be trained to identify hate speeches and controversial terms, it could act as a shield to protect certain users from being hurt. The company already claims that its AI tools report a greater number of offensive photos than humans.
Facebook’s newest technology to improve user experience sounds awesome, but what such a powerful tool could represent in terms of privacy issues remains unclear. A system that automatically analyzes text contained in every single post, video and photo may pose a threat, especially if it also scans private messages.
Last month the company was the target of a class-action lawsuit that accused it of violating federal privacy rules by looking at private messages between users of the social networking site.
— VentureBeat (@VentureBeat) June 2, 2016