To get a better browsing experience, please use Google Chrome.Download Chrome
Free TrialAsk for Price
  • Products
  • Solutions
  • Customers
  • Blog
  • API Documentation
  • About Us
  • Demo


NEXTDATA Solutions: A good manual review platform must have three characteristics

Aug 2, 2023

Content review is one of the necessary links of the information platform and content platform. If the review is too slow, it will also affect the user experience. In fact, in addition to content platforms, all platforms involving user-generated content may have content risks, such as shopping malls, payment platforms, etc. Common scenarios such as avatars, nicknames, profiles, and even secrets such as transfer information also have certain content risks.
Auditing is of course an important means of avoiding risks. Currently, content platforms on the market mainly use a combination of "machine review + human review", and few have only a single review method.
Even so, labor is still the most critical and indispensable part of it. On the one hand, algorithms will eventually have their limitations and cannot understand complex variants; on the other hand, countries also have different regulatory pressures.
So, what kind of manual review platform can be counted as a good review platform? This article will talk about NEXTDATA's thinking on this issue, as well as some practical experience.

1. A Good Review Platform Should Have the Characteristics of High Efficiency and Low Cost

The cost is reflected in the construction cost, training cost, labor cost and timeliness cost.
The first is the construction cost. It is estimated that the initial investment of a self-built manual review platform is on the order of one million. For start-up companies or middle-income companies, this is not a small amount of expenditure. External procurement can greatly save money and time costs.
The second is the cost of manual training. The learning period for new recruits is long, which may delay product launch. This can also be shortened by purchasing mature services.
The third is labor costs and timeliness costs.
The public recruitment information of Wuhan auditors published by a UGC platform shows that the monthly salary of auditors is 4,000-6,000 yuan. Calculated according to the 8% provident fund payment ratio, the labor cost is between 5344 yuan and 8016 yuan. If the team size is 2,000 people, the annual labor cost may be between 130 million and 190 million yuan.
In terms of timeliness and cost, if the system is reviewed first and issued later, the tasks will be backlogged and the user experience will be poor. If the system is issued first and reviewed later, risky content cannot be discovered in time. When a stock review is required in the face of sudden public opinion, the risk caused by manpower limitations may be irreparable.
More savings in labor and time-efficiency costs are achieved through pre-machine audits. The computer review process needs to have enough tags to identify risky content and recall, and at the same time classify and label problems in a more efficient way to improve the efficiency of human review. For example, the text risk is highlighted and the risk type is marked; the risk location and risk type are marked on the picture; the audio is divided into semantic and non-semantic for identification, and the semantic is converted into text for quick screening; video or live broadcast is divided into picture frames and audio for review.
For example, for a chat room app in a certain language, the NEXTDATA machine review + human review platform will review nicknames, avatars, personal profiles, and personal signatures according to risk levels and risk types, and realize special reviews in special areas, and increase the efficiency of single-person text/pictures from 10,000/day to 20,000/day.
In addition, the cost of timeliness also requires the real-time monitoring and dynamic adjustment of public opinion on the platform side, such as real-time dynamic updates for product compliance, Internet governance policies, Internet control and other related information, and a dedicated public opinion center to monitor hot spots in real time to ensure information synchronization at the first time.

2. It Should Have the Characteristics of Accurate Self-iteration

The disadvantages of machine review are that in terms of text, Chinese word segmentation affects machine judgment, the database update speed cannot match the speed of new words, English or symbol variants affect recognition, and machines cannot recognize complex semantics; in terms of images, it is difficult for machines to recognize the authenticity of images and cannot Understand the specific meaning contained in the image.
The problems of few data samples and misjudgment of variant identification can be solved by collecting samples to train the algorithm, or by purchasing externally in one step. In the stage where the market needs to be seized quickly, externally purchasing can greatly shorten the time spent on self-built review platform.
Questions about complex semantics and image meaning need to be handed over to human reviewers. However, the disadvantage of manual review is that the standards are different, and there are blind spots in human brain knowledge.
Through the front-end machine review, the standard can be unified as much as possible. The more detailed and accurate the label, the more it can avoid individual differences in manual review. -Map-Map of China is missing", etc. Manual review can determine whether it is a real risk content.
The auditor's knowledge is limited, and the platform needs to provide a sufficiently detailed knowledge graph. For example, if a certain audio is recognized as a speech by a religious leader, or a certain video is recognized as a national leader, the knowledge graph can help the auditor understand the detailed information. Improve review accuracy. Allowing quality inspection/re-examination of auditors and business dimensions, and defining the quality inspection/re-examination ratio according to personnel and business conditions, can further ensure content security.
More importantly, the results of manual review should feed back to the machine review model, continuously optimize the effect of machine review, and form a closed loop of effect iteration. Improve the accuracy of the pre-machine review, and further improve the overall accuracy and efficiency.

3. A Good Audit Platform Needs to be Easy to Use and Easy to Operate

For enterprises, a good audit platform can be quickly accessed through the API interface, and allows the deployment of public cloud or private cloud according to the enterprise's own situation.
For auditors, a platform that has a simple interface, clear risk classification, supports rapid risk location and visual display, and provides knowledge map references can greatly reduce auditors' operating costs.
Content review is a job of high importance and a high sense of mission in the industry. It can almost be said to be the last line of defense to maintain a "clean and clear" Internet. generate a great burden. In addition to technical characteristics, a good human review platform also needs to have a certain degree of humanistic care to improve the work experience of content reviewers. Technical means such as platform prompts and key position coding will improve the reviewer experience to a certain extent, but how? Making this irreplaceable job more warm is a topic that all enterprises and audit platforms need to constantly think about.
NEXTDATA has launched the "Tianjian Intelligent Review Platform", which integrates the advantages of AI smart review and manual review, deeply adapts to the needs of human review, and creates a warm, humanized and efficient manual review platform.

Live Chat