Hierarchical Multi-Label Classification for Knowledge Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for determining knowledge graphs struggle with efficiently classifying instances into hierarchical labels, particularly in multi-task settings with limited computational capacity, as they often require multiple transformer models and are not optimized for leveraging training data effectively.
Innovation Solution
A computer-implemented method and device that uses a hierarchical multi-label classification approach, where a sequence of tokens is embedded and classified by multiple classifiers in a tree hierarchy, with shared transformer-based neural language models and separate classifiers trained for different tasks, ensuring efficient computational use by leveraging hierarchical relations and multi-task learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple transformer models are used for hierarchical classification, then classification accuracy is improved, but computational capacity requirements increase
Solution Approach 1:
The patent combines multiple classification tasks into a single transformer model by sharing the base model across different hierarchical levels. The shared transformer processes inputs for both parent and child label classifications, reducing the number of separate models needed while maintaining classification accuracy through multi-task learning mechanisms.
Solution Approach 2:
The shared transformer model serves multiple functions by performing classification for different hierarchical levels (parent and child labels) simultaneously. This universal model structure allows a single computational engine to handle multiple classification tasks that would traditionally require separate specialized models.
2Measurement precision
If multiple separate classifiers are trained for different tasks, then task-specific performance is improved, but training data utilization efficiency decreases
Solution Approach 1:
The patent merges multiple task-specific classifiers into a unified multi-task learning framework where a single transformer model performs all classification tasks. This consolidation allows training data to be leveraged across multiple tasks simultaneously, improving data utilization efficiency while maintaining task-specific performance through separate output heads for each classification task.
Solution Approach 2:
The shared transformer model enables continuous utilization of training data across multiple tasks during the training process. Rather than training separate models on different data subsets, the unified model continuously learns from all available training data for all tasks in an integrated manner, maximizing the useful action extracted from the training dataset.
3Stability of the object's composition
If hierarchical classification is implemented with multiple levels, then label assignment consistency is improved, but device complexity increases
Solution Approach 1:
The patent merges the complexity of managing multiple hierarchical classifiers into a single unified transformer model with shared parameters. This consolidation simplifies the overall device architecture by eliminating the need to manage and coordinate multiple separate classifier systems, while the hierarchical structure itself maintains label assignment consistency through its design.
Solution Approach 2:
The universal transformer model handles multiple hierarchical classification levels through a single structure, reducing device complexity by eliminating redundant components. The model's multi-functional design allows it to process inputs and generate predictions for different hierarchical levels without requiring separate specialized structures for each level.
Data Source
AI summary
A device and a computer-implemented method of determining a knowledge graph. The method includes determining an embedding for a sequence of tokens of an instance, determining a first classification for the contextual embedding at a first classifier, determining if the first classification meets a first condition, adding to the knowledge graph a first link between a first node of the knowledge graph representing the instance and a node of the knowledge graph representing the first classification when the first classification meets the first condition and not adding the first link otherwise.