[ad_1]
Memes are the language of the web and Facebook wants to understand them better.
Over the years, Facebook AI teams have made significant advances in computer vision and natural language recognition. Today, they announced some of their latest work to combine advances in both areas. A new system, called "Rosetta," helps Facebook and Instagram teams identify text in images to better understand their subject and classify them more easily for search or to report abusive content.
They are not all the same; The tool analyzes more than a billion images and video images in multiple languages in real time, according to a company blog post.
Rosetta uses recent advances in optical character recognition (OCR) to scan an image and detect the text present. The characters are then placed in a frame that is then analyzed by convolutional neural networks that attempt to recognize the characters. and determine what is communicated.
via Facebook
This technology has been in use for a long time – Facebook has been working with OCR since 2015 – but its implementation across the company's extensive networks offers a scale of scale that has pushed the company to develop new detection and recognition strategies. characters.
If you are interested in some of the technical details of what they have done here, check out the team's research paper on the subject.
Facebook has many reasons to be interested in the text that accompanies videos or photos, especially with respect to their content moderation needs.
The identification of spam is quite simple when the textual description of a photo is "Bruh !!! Or "or" 1 like = 1 prayer ", but videos and photos using similar techniques seemed more present in calendars as Facebook tweaked its algorithm to promote" time well spent ". The same is true for hate speech, which can be much more easily shared when all email is encapsulated in an image or video, making text overlays a useful tool.
The company claims that this system presents new challenges in terms of multilingual support, as it currently uses a unified language model and most of the available training data is currently in the Latin alphabet. In the company's research paper, the team says it has put in place strategies to create new language support by reassigning existing databases.
As Facebook seeks to offload the work of moderators of human content and to allow its newsflow algorithms to sort content based on assigned classifications, a tool like this has a lot of potential for you.
Source link
