Dealing with controversial content including terrorist videos, hate speech, porn or violence has been a constant issue for various social media sites including Facebook, Twitter and YouTube. But Facebook, on the final day of the F8 2018 developers conference, announced that it is using technologies such as artificial intelligence, computer vision and machine learning to remove “bad stuff” even before its reported.
Facebook’s VP of product management, Guy Rosen, in a post mentioned that these technologies can help identify possible bad content. For instance, someone expressing a thought of suicide can be helped by human expertise after such a post gets detected by the technology. Rosen mentioned that Facebook removed 99 percent of the almost two million pieces of ISIS and al-Qaeda-related content posted on the platform in Q1 2018.
The new software is constantly improved by being fed with data to identify nudity, graphic violence and hate speech. The hate speech detection is available only in English and Portuguese at the moment.
The software is also helping in identifying fake accounts, spam, terrorist propaganda and suicide prevention. He mentions that closing fake accounts helps in fighting spam, fake news, misinformation and bad ads.
Rosen also mentioned that technology such as artificial intelligence is not making quick progress as it’s not able to identify the context. Someone talking about their own drug addiction or encouraging others to do drugs, differ in context and need human review. Also, there have been cases in the past where algorithms used for Facebook’s Trending Topics, could not differentiate between news and satire.
Other problems which come with technology includes limited data in non-English languages. The AI needs a large amount of data to detect behavioural patterns which is not easily available for less widely used languages.
But, according to Rosen, Facebook is investing in technology to improve its efficiency and reach.
Facebook AI Research team is working on a new area called as multi-lingual embedding to overcome the language-barrier challenge and increase its accuracy across new languages.