Improving Content Moderation with Artificial Intelligence
In recent years, live video has gained popularity and a large number of live broadcasting platforms have mushroomed. However, this has also brought enormous challenges to the supervision of live content. Meanwhile, the country’s requirements for live content supervision have become increasingly stringent. Unlike text content, multimedia content is harder to be reviewed. The resource consumption and the technical threshold are also relatively higher. This has become a bottleneck restricting the healthy development of live broadcasting services, especially for some small and medium-sized live broadcasting platforms that have just started.
For content moderation, many platforms still hire a large number of human content reviewers to review massive streams of videos and find videos that are not compliant. However, due to the massive amounts of content, this method often ends up with inconsistent standards and results. Moreover, this method of content review is not scalable and can be costly when employed at a large scale.
Fortunately, artificial intelligence (AI) is now offering a way forward; AI pre-review + manual review is becoming a norm for such reviews.
Traditional AI-Powered Solutions
In a typical approach for AI-powered video content review, enterprises would set up their own AI teams to train their detection models based on their service scenarios and resources. The advantage of this method is that they can better control their content filter based on actual needs. However, this method also has some disadvantages, for example:
- Users need to capture frames by themselves; this simple work may consume a lot of computing resources.
- Costs for massive image storage and transmission.
- Limited by the technical ability and manpower; the accuracy of the review algorithms is not satisfactory.
Live Broadcasting Review Solution
Alibaba Cloud ApsaraVideo Live provides a content review feature that can solve the above challenges convincingly. ApsaraVideo Live lowers the technology adoption barrier; users can quickly enable the live broadcasting review service on the console. This solves small and medium live broadcasting platforms’ content moderation pain points.
Live Broadcasting Review Architecture
- A broadcaster streams to the live broadcasting center and triggers the content review task
- The system captures the Frames in the live stream at a specified rate
- Next, it performs a comprehensive content moderation on the captured frames in real-time using the content moderation engine
- It sends the results to the control service for comprehensive judgment on whether or not it is necessary to notify the customer and save the evidence frames
Live Broadcasting Review Features
It is transparent for users, and users can easily enable the live video review service on the console without paying for extra storage, traffic, and computing power. Users can perform a quick review and block problematic or suspect videos through self-built platforms or our live broadcasting review platform.
Powerful Algorithm Support
Organization can quickly and easily introduce advanced algorithms that rely on Alibaba Cloud’s accumulated video and AI technologies. The industry-leading live broadcasting review algorithms enable comprehensive multi-frame determination, which helps to reduce the misjudgment rate.
- Inappropriate content filter
- Low-quality live video detection (for frequent live streams without meaningful content, or live streams that have ended)
Detection Model Customization
Due to the diversity of users’ needs and application scenarios, the live broadcasting review provides a more flexible solution where a user can dynamically utilize a different review policy for each stream. The algorithm model can also be different depending on the streams; this is very necessary because different scenarios vary significantly. In practice, we have found that the misjudgment rate is high for some live streams.
For example, typical live video review algorithms often misinterpret boxing matches as inappropriate content because it involves boxers who are shirtless.
Another example is video games. Because the scenario in a game is rendered, the texture is different from a normal image captured by a camera. Moreover, a lot of popular games have a violent backdrop. Therefore, correctly distinguishing between a game scene and a real scene is also vital; for example, systems usually misjudge many scenes in a game as terrorism.
Therefore, different streams must be subdivided according to the scenario via algorithms as this can improve accuracy levels and make the live review more practical. Alibaba Cloud’s live video review solution provides stream-level fine algorithm control to cover these scenarios better.
Organizations can use different detection scenarios and different frame capture rates for different streams dynamically. The frame capture rate can be up to 1 fps. Further, it is possible to apply different monitoring strengths for different video streams and monitor high-risk streams intensively. All this can further reduce costs.
Practical Auxiliary Functions
In practice, several independent images may not be sufficient as evidence to identify violations. In combination with the real-time recording function of the live broadcasting platform, organizations can record problematic streams in real time to collect evidence for manual review.
For some complex scenarios, such as those involving political content, major risks may occur in audio instead of video. To avoid being caught by the video filters, content providers may sometimes include some indecent content in speech format. Organizations can use Alibaba Cloud speech recognition system for audio content, enabling a multi-dimensional live broadcasting review.
Alibaba Cloud ApsaraVideo Live review solution can easily provide users with live review capabilities, significantly reduce the cost of human resources, and improve review efficiency. To learn more about ApsaraVideo Live, visit https://www.alibabacloud.com/product/apsaravideo-for-live