Naver has advanced its artificial intelligence (AI)-based comment detection system by one step to reduce malicious comments.
Naver said on the 29th that it applied the upgrade for "AI Cleanbot 3.0" starting on the 29th. The core of this revamp is to focus on filtering expressions that devalue life and secondary victimizing comments directed at victims and bereaved families of incidents and accidents. Going beyond simply blocking profanity or slurs, it analyzes the article content to which the comment is attached to judge malicious intent.
AI Cleanbot 3.0 does not look only at the comment sentence in isolation but grasps the context by combining the article headline and body text. It said the model has been improved to more precisely judge whether the same expression was used to mock, hate, or belittle depending on the article's content. Naver plans to use this to reduce the exposure of comments that cause additional harm to victims and bereaved families in articles related to social incidents and accidents.
Since first introducing AI Cleanbot in 2019, Naver has steadily expanded the detection range. At first, it was close to keyword detection focused on profanity and slang, but since 2020 it has been improved to detect insulting expressions without explicit profanity by reflecting sentence context. It has since strengthened responses to sexually discomforting expressions, hate and discriminatory expressions, and malicious comments that circumvent detection using symbols and characters.
Policy measures are also being carried out in parallel. For articles where malicious comments exceed a certain threshold, Naver automatically disables the comment service and shows users a guidance message along with a "Green Internet" campaign banner. It has also moved to improve the news comment environment by restricting the operation of comments on political and election-related articles.
Kim Su-hyang, a Naver leader, said, "We are continuously strengthening Cleanbot's performance to respond to newly emerging expressions of hate, belittlement, and discrimination," and added, "We will reflect diverse opinions to build a healthy comment culture."