This article is from "TechWeb"
In the era of "everyone is an anchor and everything can be broadcast live", in order to strictly prevent the emergence of vulgar content, China's leading game live broadcast platformDouyu (DOYU.US)Continue to invest in technology research and development, optimize and improve platform rules, crack down on vulgar content, and build content firewalls with AI.
Since March this year, Douyu has conducted a classified inspection and rectification of the live broadcast content, conducted a comprehensive and in-depth investigation of six categories of malicious marketing accounts, including pornography, rumors, infringement and counterfeiting, and the title party, and dealt with the problem accounts seriously and quickly. In June this year, Douyu cleaned up a total of 4099 problem accounts, including 3200 in low-level vulgarity, 861 in title party, 20 in rumor mongering and spreading rumors, and 18 in infringement and counterfeiting. Up to now, Douyu content improvement measures have achieved practical results.
Under Douyu's content supervision system, all broadcasters on the platform need to go through three firewalls of "real name verification + face recognition + manual audit", which can not only prevent bad anchors from broadcasting from the source, but also prevent minors from broadcasting live broadcasts. And can track individuals as soon as the anchors break the rules to prevent the occurrence of subsequent vicious events.
After the VJ opens the live room, the live video content produced by the VJ keeps running in the AI system, and each picture will be identified through the program. Once the machine automatically identifies bad content such as vulgarity, parody, abuse and other bad content, and intercepts illegal images and videos, the platform can achieve a "second pass" for these bad content. In the face, outdoor and other pan-entertainment sections, the system will also synchronously capture cross-transmitted sound information for slicing, analyze through the data model, and intelligently judge and identify through key points such as voice, semantics, on-screen text, and so on. Filter the content of the problem and push it to manual review.
It is understood that Douyu's AI audit technology has developed quite mature. at present, in the live broadcast field, Douyu has very high porn detection accuracy for pictures, video and audio, and has been able to export this technical capability. Compared with other platforms, Douyu has larger and unique user data and behavior data, and is able to analyze and calculate based on massive user data. Compared with the early Douyu can only use third-party data for image recognition, the current platform has achieved independent modeling. Compared with the third-party data, after long-term data screening, cultivation and self-learning, Douyu risk identification system has accumulated a large number of violation characteristic data and built a strong computing power, which makes its own risk identification model more accurate.
In addition to automatic supervision, Douyu adopts a combination of machine audit and manual audit to monitor all live channels 24 hours a day through background monitoring, personnel inspection, machine identification and report processing. In order to further prevent "missing fish", Douyu also set up functions such as phone reports, web pages, and one-click APP reports, especially gathering and recruiting enthusiastic viewers to supervise and inspect live broadcast-related content across the platform. Up to now, the total number of participants in Douyu's patrol team is more than 170000, and more than 2000 enthusiastic viewers of Douyu have joined.
In addition, Douyu has specially formulated a sound emergency handling system, covering the "emergency handling system for public security emergencies, emergency response plans, and Internet emergency response plans." improve the emergency handling capacity of major security incidents and emergency sensitive events, minimize the impact of events from the root causes, and minimize the damage and losses caused.
Under Douyu's strict examination and ban measures, the "Qinglang" operation has been quite effective. Douyu said that in the future, the platform will continue to improve the content review mechanism, enrich the content review team, strictly implement the requirements of 24-hour monitoring, and continue to promote the ecological health of the webcast industry.
(editor: Li Guojian)