Naver's AI for Blocking Malicious Comments Knows... "The Word 'Sseuregi' is OK, the Insult 'Sseuregi' is NO"

[Asia Economy Reporter Buaeri] It has been 17 years since Naver started its comment service. The comment culture, which once served as a 'public forum,' has deteriorated into a social problem causing abusive language and violence. As the situation worsened, Naver temporarily suspended comments on entertainment and sports articles. Recently, Naver has been waging a war against malicious comments (malcomments) by introducing artificial intelligence (AI) to restore the 'comments' space as a place for communication.

Cleaning an Average of 30,000 Malcomments Daily

To address the malcomment issue, Naver applied the AI system 'Cleanbot' to all news comment services in November last year. Lee Gyuho, the leader of Naver's User Feedback Platform who developed Cleanbot, explained in an interview with Asia Economy on the 20th, "Cleanbot is a malcomment cleaner." Cleanbot detects abusive or insulting expressions in news article comments and then 'hides' them.

Lee said, "Even a single malcomment can hurt someone, cause fights, and trigger conflicts between groups," adding, "Malcomments blur the function of comments as a communication space. Through Cleanbot, we made it possible for comments to function properly as originally intended." Naver is gradually expanding the scope of Cleanbot's application. Initially introduced only in sports and webtoon sections, it was expanded to all news comments, and recently, Cleanbot has been cleaning malcomments on 'Naver TV' and the music platform 'Vibe' as well.

Since the introduction of Cleanbot, reports of malicious comments have decreased by about 19%. Lee stated, "Although there are fluctuations depending on the issue at hand, 15% to 20% of all comments written are filtered out." According to Naver Data Lab, as of the third week of November, an average of 200,000 to 300,000 comments are posted daily. Based on this, Cleanbot is catching more than 30,000 malcomments daily on average.

Some voices emphasize 'freedom of expression.' Regarding this, Lee said, "We are always concerned about conflicts with freedom of expression and strive to delicately balance it," emphasizing, "While standards for malcomments may vary from person to person, Cleanbot's filtering criteria are at a level that most people can agree upon."

Lee Gyu-ho (right), User Feedback Platform Leader, and Hwang Tae-hyun, Manager.

AI That Detects 'Verbal Sexual Harassment' Also Released

Naver has steadily advanced Cleanbot AI technology over the past year. It continuously trained the AI on data such as types of malcomments and sentence patterns. The 'Cleanbot 2.0,' introduced last June, has the ability to judge malcomments by understanding the entire context. For example, the word 'trash' was not filtered as a malcomment before because the AI only detected abusive keywords. However, now Cleanbot detects and catches malcomments with violent meanings, such as in the phrase 'you trash-like human.' Expressions written with spacing tricks to avoid being blocked by Cleanbot are also caught. Naver evaluates Cleanbot's accuracy to be around 95%.

If AI continues to develop, there is a possibility that the previously closed entertainment and sports news comment sections may reopen. Lee said, "AI cannot catch 100% of malcomments. We decided to temporarily suspend comments because even a single comment can be fatal to the person involved," adding, "If we develop good technology, we might consider reopening someday."

Naver plans to introduce 'Cleanbot Red' (tentative name) next year, which detects sexual harassment and verbal sexual violence. The goal is to precisely catch expressions that clearly constitute sexual violence among various types of malcomments. Through this, Naver hopes that internet comment sections will return to their original role as 'forums for communication.' Lee said, "I believe criticism and condemnation are different," and added, "Our goal is to create a space where constructive criticism can be made and where people can support each other."

This content was produced with the assistance of AI translation services.