iTAG allows you to use intelligent labeling configurations to perform data pre-labeling. After you complete data pre-labeling, you can perform formal data labeling based on the pre-labeling results. In formal data labeling, all you need to do is verify and correct the pre-labeling results. This improves labeling efficiency. iTAG supports offline and online pre-labeling. In offline pre-labeling, you must upload data in an acceptable format for automated pre-labeling. In online pre-labeling, the system uses an API operation to automate data pre-labeling. This topic describes the format of offline pre-labeling datasets and the considerations for creating pre-labeling jobs. This topic also provides examples of pre-labeling results.
Limits
iTAG pre-labeling supports only classification scenarios, such as image or text classification.
Offline pre-labeling
Formats of datasets used for offline pre-labeling
Prepare the offline pre-labeling file named prelabel_offline.manifest and create a dataset in an Object Storage Service (OSS) bucket from the offline pre-labeling file.
The following code block shows the format of the offline pre-labeling dataset. The dataset must contain custom fields and the source field. You can specify one or more custom fields in the offline pre-labeling dataset.
{ "data": { "label": "label_2", "source": "Alibaba Group's 10th anniversary and the founding of Alibaba Cloud Computing" } }Parameters:
label: the pre-labeling result that is generated offline.
source: the source data.
For more information about the format requirements of different types of files used for labeling, see Create a dataset for a labeling job. If your offline pre-labeling dataset does not meet the format requirements, the pre-labeling results may not be displayed as normal.
For more information about how to create a dataset in an OSS bucket from an offline pre-labeling file, see Create and manage datasets.
Considerations for creating offline pre-labeling jobs
When you create a labeling job, you can configure the following parameters on the Intelligent Labeling Configurations wizard page.
In the Service Configurations section, you can configure the following parameters.
Parameter
Description
Labeling Method
Available options:
Do Not Use: If you select this option, intelligent labeling is not used.
Use Offline Prelabeling Result: If the dataset that you uploaded contains pre-labeling results, you can select this option to display the pre-labeling results in the iTAG console.
Online Service Prediction: If the dataset that you uploaded does not contain pre-labeling results, you can select this option to call an online prediction service to process the dataset. The prediction results are displayed as pre-labeling results in the iTAG console.
Mapping between Predicted Result and Topic
Map the pre-labeling result column in the offline pre-labeling dataset to a topic. This way, you can specify the pre-labeling results of different topics.
If you want to label data for multiple topics, click Add Mapping between Predicted Result and Topic to add more mappings.
In the Service Configuration section, you can select one of the following options for Effective Process:
Prelabeling: If you select this option, the pre-labeling results are effective only in the pre-labeling stage. The workers must verify and correct the pre-labeling results in the formal data labeling stage.
Formal Labeling: If you select this option, the pre-labeling results are effective in the formal data labeling stage. Workers do not need to label data in the formal data labeling stage. They can proceed to the verification or acceptance stage.
Examples of offline pre-labeling results
The pre-labeling results of different topics are displayed on the labeling results page, as shown in the following figure.
Online service prediction
iTAG supports online service prediction. Before you use iTAG to perform online prediction, you must create a model service. For more information, see Model services.
Select Service: selects a model service.
Service Input Parameter Mappings: specifies mappings between input parameters and data fields that need to be recognized.
Mapping between Predicted Result and Topic: specifies mappings between topics and output fields of the model service.
