Hierarchical Intent Labeling for Consistent ML Training Data

The patent segments the large intent space into a hierarchical structure with parent intents at higher levels and child intents at lower levels. Labelers first select from parent intents, then progressively refine to child intents. This segmentation reduces the number of options presented at each decision point, eliminating choice overload while maintaining comprehensive intent coverage across the entire hierarchy.

The patent introduces a hierarchical dimension to the intent selection process. Instead of presenting all intents in a single flat list, the system organizes intents across multiple hierarchical levels (parent, child, grandchild intents). This dimensional transformation allows labelers to navigate through layers of abstraction, reducing cognitive load while preserving the ability to select from a comprehensive set of intents.

The patent segments the complete set of intent labels into hierarchical groups, presenting only the relevant subset at each level. Labelers first see parent intents, then progressively see child intents only after selecting a parent. This segmentation ensures that the complete intent space remains available for coverage while preventing choice overload by limiting simultaneous options to manageable subsets.

The patent implements preliminary action by requiring labelers to first select parent intents before accessing child intent options. This preliminary selection step filters the available intent space, preparing a narrowed-down set of relevant options for the next labeling decision. This sequential approach improves identification accuracy by eliminating irrelevant options before the labeler makes the final intent selection.

The patent implements dynamics by allowing the intent label ordering to adapt based on context. During the labeling process, intents are presented in hierarchical order for systematic selection. However, when generating the final intent label output, the system dynamically reorders the components to reflect natural language conventions and improve readability, rather than strictly maintaining the hierarchical presentation order.

The patent applies local quality by treating different stages of the labeling process differently. In the selection phase, hierarchical ordering provides systematic structure. In the output phase, natural language ordering provides readability. Each stage has its own optimal ordering strategy applied locally, rather than imposing a single global ordering constraint throughout the entire process.