Skip to main content
Version: 0.91

Hierarchical configuration for classification and extraction tasks

In this page, you will learn how to create a hierarchical task (classification, extraction, etc.).

We will demonstrate how to use Snorkel Flow for hierarchical classification problems by building upon the application templates described in the Document classification: Classifying contract types tutorial.

Create a dataset and application​

Follow the dataset and application creation from the Document classification: Classifying contract types tutorial.

Create hierarchical configuration​

After loading the datasources and creating the application template, we can create the hierarchical configuration by:

  • Selecting the dropdown bar at the model node (the three-dot next in the model node)
  • Select Build hierarchical model. This selection will prompt an interface for inputting sub-labels
  • Click Split into submodels once you input all sub-labels
  • Snorkel Flow will create a hierarchical DAG for the appropriate template

For example, we can split “services” into 2 sub-classes “management-services” and “non-management-services”; splitting “employment” into 2 sub-classes “executive-employment” and “non-executive-employment”. The two classes “loan” and “stock” can stay as they are.

hierarchy_11.webp

The expected output will look as follows:

hierarchy_21.webp

About the DAG:

  • The Model nodes are where we iteratively write labeling functions, train the models, and analyze performance.
  • The LabelFilter and ColumnFiller will ensure Snorkel Flow passes the predictions into the correct models.
  • The ConcatRows will aggregate the predictions across all models and create the resulting dataframe. This node is also the last node of the hierarchical task.

Iteration and development​

The iteration process will happen at every Model node. Each model will perform a classification/extraction/etc task for each sub-labels group (e.g., “management-services” and “non-management-services”, “executive-employment” and “non-executive-employment”). You can upload the corresponding ground truths, write labeling functions, train models, and analyze performance.

Final predictions​

The ConcatRows node will output a dataframe, with the preds_str column containing the final prediction.