Platform For AI:Configure intelligent pre-labeling

Offline pre-labeling

File format

Prepare a manifest file, such as prelabel_offline.manifest, and use it to create an OSS dataset.

• Each manifest entry must contain a source field and at least one custom field for the pre-labeled result. Multiple custom fields are supported.

  {
      "data": {
          "label": "label_2",
          "source": "Alibaba Group celebrates its 10th anniversary and establishes Alibaba Cloud"
      }
  }

  Fields:

  • label: Pre-labeled result generated offline.

  • source: Original data to label.

  For format requirements by labeling type, see Create a dataset for a labeling job. Incorrectly formatted manifest files cause pre-labeled results to display incorrectly.

• Create the dataset from this manifest file as an OSS dataset. For more information, see Create and manage datasets.

Job configuration

When creating a labeling job, configure the parameters on the Intelligent Labeling Configurations wizard page.

• In the Service Configurations section, configure the parameters described in the following table.

  Parameter	Description
  Labeling Method	Labeling method: Do Not Use: Disables intelligent labeling. Use Offline Prelabeling Result: The dataset already contains pre-labeled results. iTAG displays the results for workers to review. Online Service Prediction: The dataset does not contain pre-labeled results. iTAG calls an online prediction service to generate them.
  Mapping between Predicted Result and Topic	Maps pre-labeled result columns in the manifest file to topic names. For multiple topics, click + Add Mapping between Predicted Result and Topic to add more column-to-topic mappings.

• In the Service Configuration section, set the Effective Process parameter:

  • Prelabeling: Pre-labeled results apply only to the pre-labeling stage. Workers review and correct the results during formal labeling.

  • Formal Labeling: Pre-labeled results apply directly to formal labeling. Workers skip manual labeling and proceed to verification or acceptance.

Result example

Each question on the labeling page displays its pre-labeled result.