Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Augmenting Knowledge Graph Hierarchies Using Neural Transformers

Sanat Sharma, Mayank Poddar, Jayant Kumar, Kosta Blank, Tracy King

TL;DR

The problem addressed is the flat, limited hierarchies in domain-specific knowledge graphs, which hinder semantic organization and downstream recommendations. The authors propose a transformer-based pipeline that defines top-level $L_1$ categories, uses few-shot prompting for node-to-category classification, and expands the hierarchy with a generator module employing either cyclical or one-shot strategies, supported by an auto-ingest workflow. Key contributions include achieving 98% coverage for intents and 99% for colors, a comparative analysis of generation strategies, and a human-in-the-loop evaluation confirming relevance and correctness. The findings demonstrate a production-ready method to enrich KG hierarchies, improving navigation, semantic relationships, and SEO-related connections, with Adobe Express color pages as an early user-facing application.

Abstract

Knowledge graphs are useful tools to organize, recommend and sort data. Hierarchies in knowledge graphs provide significant benefit in improving understanding and compartmentalization of the data within a knowledge graph. This work leverages large language models to generate and augment hierarchies in an existing knowledge graph. For small (<100,000 node) domain-specific KGs, we find that a combination of few-shot prompting with one-shot generation works well, while larger KG may require cyclical generation. We present techniques for augmenting hierarchies, which led to coverage increase by 98% for intents and 99% for colors in our knowledge graph.

Augmenting Knowledge Graph Hierarchies Using Neural Transformers

TL;DR

The problem addressed is the flat, limited hierarchies in domain-specific knowledge graphs, which hinder semantic organization and downstream recommendations. The authors propose a transformer-based pipeline that defines top-level categories, uses few-shot prompting for node-to-category classification, and expands the hierarchy with a generator module employing either cyclical or one-shot strategies, supported by an auto-ingest workflow. Key contributions include achieving 98% coverage for intents and 99% for colors, a comparative analysis of generation strategies, and a human-in-the-loop evaluation confirming relevance and correctness. The findings demonstrate a production-ready method to enrich KG hierarchies, improving navigation, semantic relationships, and SEO-related connections, with Adobe Express color pages as an early user-facing application.

Abstract

Knowledge graphs are useful tools to organize, recommend and sort data. Hierarchies in knowledge graphs provide significant benefit in improving understanding and compartmentalization of the data within a knowledge graph. This work leverages large language models to generate and augment hierarchies in an existing knowledge graph. For small (<100,000 node) domain-specific KGs, we find that a combination of few-shot prompting with one-shot generation works well, while larger KG may require cyclical generation. We present techniques for augmenting hierarchies, which led to coverage increase by 98% for intents and 99% for colors in our knowledge graph.
Paper Structure (6 sections, 8 figures, 1 table)

This paper contains 6 sections, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Knowledge Graph Work
  3. Approach
  4. Few-Shot Prompting
  5. Generation Module
  6. Hierarchical KG Evaluation and Conclusions

Figures (8)

  • Figure 1: Adobe Express SEO page for birthday with related pages powered by the intent-based KG.
  • Figure 2: Hierarchies allow us to understand and compartmentalize nodes
  • Figure 3: Hierarchy Generation Approach
  • Figure 4: Cyclical Generation requires classification of nodes and addition at each level
  • Figure 5: LLMs can sometimes confuse similar semantic meaning with parent-child relationship
  • ...and 3 more figures