[Paper Review] Knowledge Graphs on the Web -- an Overview

Nicolas Heist, Sven Hertling+2 more|arXiv (Cornell University)|Mar 2, 2020

Advanced Graph Neural Networks42 references38 citations

TL;DR

The paper surveys publicly available cross-domain knowledge graphs, comparing size, content, schema, and inter-graph overlap to guide KG selection for applications including explainable AI.

ABSTRACT

Knowledge Graphs are an emerging form of knowledge representation. While Google coined the term Knowledge Graph first and promoted it as a means to improve their search results, they are used in many applications today. In a knowledge graph, entities in the real world and/or a business domain (e.g., people, places, or events) are represented as nodes, which are connected by edges representing the relations between those entities. While companies such as Google, Microsoft, and Facebook have their own, non-public knowledge graphs, there is also a larger body of publicly available knowledge graphs, such as DBpedia or Wikidata. In this chapter, we provide an overview and comparison of those publicly available knowledge graphs, and give insights into their contents, size, coverage, and overlap.

Motivation & Objective

• Motivate the use of knowledge graphs as a versatile knowledge representation for integrating heterogeneous data.
• Provide an overview of publicly available cross-domain knowledge graphs (e.g., DBpedia, Wikidata, YAGO, BabelNet, NELL, CaLiGraph, Voldemort).
• Characterize graphs by contents, size, schema complexity, and overlap to aid application-specific KG selection.
• Discuss creation techniques (manual, structured data, unstructured data) and their implications for coverage and quality.

Proposed method

• Classify creation methods into manual curation, (semi)structured data extraction, and unstructured data extraction.
• Describe representative pipelines: Wikipedia infoboxes mapping to ontologies (DBpedia), category/WordNet-based typing (YAGO), multilingual integration (Wikidata, BabelNet), and extraction from web data (CaLiGraph, VoldemortKG).
• Summarize general metrics for KG evaluation (instances, assertions, linking degree, classes, relations, ontology depth/width, complexity).
• Present a cross-KG comparison of size, detail, and schema complexity using open KGs such as DBpedia, YAGO, Wikidata, BabelNet, NELL, CaLiGraph, Voldemort, and Cyc.
• Explain interlinking and overlap estimation between KGs using heuristic link discovery and recall/precision-based estimates.

Experimental results

Research questions

• RQ1 What are the main publicly available cross-domain knowledge graphs on the Web and how do they differ in size, contents, and schema?
• RQ2 How do creation methods (manual, structured, unstructured) influence the coverage and quality of these knowledge graphs?
• RQ3 What is the overlap and interlinking between major knowledge graphs, and how does combining them potentially improve coverage?
• RQ4 How can metrics at instance and schema levels inform the selection of a KG for specific applications (e.g., XAI)?

Key findings

• Knowledge graphs vary by orders of magnitude in size (from tens of thousands to tens of millions of instances) and in the number of assertions (from hundreds of thousands to hundreds of millions).
• Schemas differ in depth and complexity, with DBpedia/NELL having smaller schemas, while Wikidata/BabelNet have deeper, more detailed taxonomies; Cyc shows the highest complexity due to its formal ontology.
• Wikidata generally has the largest number of instances for many classes, while YAGO often provides the most detailed (highly connected) class-level information.
• Interlinking across KGs is common and mostly complete among Wikidata, DBpedia, YAGO, and CaLiGraph, but weaker for NELL, OpenCyc, and Voldemort, indicating that multi-graph integration requires interlinking effort.
• Combining KGs can yield substantial gains in coverage, especially for complementary sources like BabelNet and CaLiGraph with DBpedia, but overlap varies by class and source.