Hierarchical Graph Embedding for Inductive Content Tagging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional content tagging systems rely solely on graph-based approaches that consider relationships between content instances, neglecting the content itself, leading to poor tagging performance and the need for frequent re-training, which is computationally expensive.
Innovation Solution
A hierarchical unified graph embedding framework that combines a language model and a graph model for co-training, using semantic embeddings to enhance graph labeling, allowing for inductive inference without re-training on new content.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional graph-based approaches are used for content tagging, then the system can leverage known relationships between content instances, but the tagging accuracy deteriorates because the content itself is neglected
Solution Approach 1:
The patent combines graph-based relationship modeling with language model-based semantic understanding into a unified framework. The graph model captures structural relationships between content instances while the language model processes the actual content semantics, and their outputs are merged to produce accurate tags that leverage both relationship and semantic information.
Solution Approach 2:
The system creates a composite approach by integrating two different modeling paradigms: the graph model that excels at capturing relational structures and the language model that excels at understanding content semantics. This composite methodology allows the system to benefit from the strengths of both approaches simultaneously.
2Reliability
If graph-based systems are re-trained frequently to maintain accuracy, then tagging performance can be maintained, but computational cost and operational expenses increase
Solution Approach 1:
The system performs preliminary action by pre-processing content through the language model to generate semantic embeddings before graph-based relationship analysis. This preliminary semantic understanding is cached and reused during tagging operations, avoiding the need for frequent re-training of the entire system while maintaining accurate tagging performance.
3Device complexity
If solely graph relationships are considered for tagging, then the system structure remains simple, but the tagging precision deteriorates due to lack of content understanding
Solution Approach 1:
The patent segments the tagging system into two distinct functional components: a graph model segment that handles relationship analysis and a language model segment that handles semantic understanding. This segmentation allows each component to specialize in its strength while maintaining a relatively simple overall structure, avoiding the need for a single complex monolithic model.
Data Source
AI summary
Approaches in accordance with various embodiments provide unified training and inference frameworks useful for co-training, and performing inferencing using, models such as language and graphing models. The frameworks can be hierarchical, in that an inner layer can include a language model for generating semantic embeddings from input text content, and these semantic embeddings can be passed as input to an outer layer that can generate graph labels based, at least in part, upon a knowledge graph having these semantic embeddings substituted for textual content. Nodes of a knowledge graph that are determined to be related to the input content can then be used for various purposes, such as to generate tags for the content or determine how to route that content.