OCs cut toxicity in comments nearly in half with the help of AI

"Odnoklassniki" began to temporarily restrict users who write offensive comments. The development based on machine learning has reduced the share and number of negative statements under posts, photos and videos by 47%.

"Odnoklassniki" implemented a machine learning model that allows you to automatically identify toxic comments and limit their authors in writing new ones for a certain period of time - from one hour to 24 hours, depending on how often the user leaves negative statements about others.

According to the OC study, negative posts make more than 40% of users who participated in the survey angry. The ML model was trained on 400 thousand statements that could be described as unfriendly or provoking negative discussion. To find out which comments OC users think are toxic and unwanted, the social network conducted a survey among its audience. The launch of the ML model has reduced the processing time for complaints about toxic posts by three times and helped people with solutions to their problems faster.

"We see from user requests that our audience has a need for a space for safe and respectful communication: one-third of all complaints on the platform are related to comments users feel are negative, and 80% of support calls about comments are targeted against 'toxicity.' We've been training this model for over a year and will continue to refine it going forward. Together with other product solutions that we have implemented over the past two years, limiting toxic commenters will allow the social network not only to return to the platform more quickly users who rarely visit OK, but also to attract new ones," - comments Alexander Moskvichev, product director of OK.

Earlier this year, the social network began to automatically detect on the platform foul words and hide them, as well as blur images of erotic nature. OK uses machine learning-based moderation to limit other unwanted and inappropriate content. For example, they block what is intended for an 18+ audience, blur and remove photos with violent scenes. Last year, the social network also launched the ML model, which restricts intrusive dating and communication with unwanted interlocutors on the platform.