To continuously improve the moderation performance, you can enable data labeling and reflux to help the model specifically learn about your moderation standard and increase identification accuracy.
If you have no moderation platform, you can use our self-service moderation platform for data moderation and reflux. We will inform you of the moderation results through the callback interface.
Log on to the Content Moderation Human Review Platform to view the image identification results in real time, and label and feedback the samples that are incorrectly identified during moderation. If you have a limited moderation workforce, you are recommended to focus on the samples that are identified as illegal or suspicious in moderation.
The labeling rule is as follows: If you think a sample is normal but it is identified as illegal or suspicious, label the sample as “Normal”. If you think a sample needs to be controlled but it is identified as suspicious or normal, label the sample as “Illegal”. After labeling a sample, click to submit it. You can also select multiple samples and submit them in a batch. The moderation results of this sample and its similar samples will be rectified in real time based on your labels, and the callback interface will send a notification to you.
Human review result notification: When you moderate the results returned by the Content Moderation API, you can set a callback action and an HTTP(S) interface. Then, after you submit a moderation action, we will use this interface to send the moderation result and the original system moderation content to you.Choose Console Settings > Content Moderation API > Self-service Moderation Result Notification to add a callback link. You can add only one callback link, and then the system automatically generates a callback seed.The callback link must support the POST mode, employs the UTF-8 encoding format, and supports two form parameters, checksum and content. The system sets values for checksum and content according to the following generation rule and format, and calls your callback interface to return the moderation result.Convention: If your server receives the result notification and returns a “200” HTTP status code, the notification is regarded as successful. If your server returns another HTTP status code, the notification is regarded as failed. The notification is retried for a maximum of 16 times.Generation rules and formats of callback result parameters:
|checksum||String||A string in the format of “uid + seed + content” that is concatenated by the SHA256 algorithm. To prevent data tampering, you can (not mandatorily) generate a string based on this algorithm when receiving a result, and verify the string against checksum.|
|content||String||JSON string format. You must resolve and reverse the string into a JSON object. The content format is described as follows:|
NOTE: If you forget your uid (account ID), retrieve it on the Alibaba Cloud console.
The content parameter contains two parts. One is the scan result (scanResult), and the descriptions of its fields and format are the same as those of the returned API call result. The other is the moderation result (auditResult), where “suggestion” indicates whether the moderation result is illegal (“block”) or normal (“pass”), and “labels” offers detailed reasons why the content is illegal.
"labels":["porn", "ad", "terrorism"]
If you have your own moderation platform, you can connect the platform to the feedback interface and feed back to us the samples that are incorrectly identified in moderation. When developing a new model, we will include your feedback data to better train the model. For more information about the feedback interface, see the interface documentation. Note that model training must be based on sufficient samples and may not take effect immediately. When necessary, you can enable the real-time data reflux function to automatically add your feedback data to the custom image library and rectify the moderation results immediately.
- At first, you must open a ticket or contact your business manager to ask the content security operation personnel to enable the real-time data reflux function for your custom image library and select a data reflux scenario. Then the system automatically creates a reflux image library for this scenario in your custom image library, and configures a blacklist and a whitelist.
- By using the feedback interface, set the “label” field of the samples that you consider as normal to “normal”, and these samples are automatically added to the whitelist, and set the “label” field of the samples that you consider as illegal to a risky value, such as “porn”, “ad”, and “terrorism”, and these samples are automatically added to the blacklist.
- You can manage a reflux image library on the console in the same way as you manage other custom image libraries, except that you cannot create or delete the reflux image library.