Knowledge Harvesting: Achievements and Challenges

Gerhard Weikum, Johannes Hoffart, Fabian M. SuchanekComputing and Software Science 2019 · OpenReview

Abstract

This article gives an overview on knowledge harvesting: automatically constructing large high-quality knowledge bases from Internet sources. The first part reviews key principles and best-practice methods. The second part points out open challenges for future research.

Introduction

For decades, a "holy grail" of computer science has been equipping machines with a structured understanding of the world. While early attempts like Cyc or WordNet relied on painstaking manual curation, they could not scale to the breadth of the modern Internet. Knowledge harvesting has emerged as the transformative process of automatically turning messy Internet data into structured, high-quality knowledge bases (KBs). By extracting entities, classes, and relationships, this technology enables machines to distinguish between "jobs at Apple" as a career search versus a fruit-picking inquiry.

Methodological Principles

Building modern KBs requires a layered approach, moving from high-confidence semi-structured data to noisier unstructured text.

Model-Driven Harvesting and Pattern-Fact Duality

Most successful KBs utilize a pre-defined schema of predicates (e.g., bornIn, foundedBy). Through "distant supervision," systems start with a few "seed facts" to identify linguistic patterns in text. These patterns are then used to discover millions of new facts across the web.

Examples of structured knowledge triples for Steve Jobs, showing types, subclasses, and relationships.

Canonicalization and Knowledge Cleaning

A crucial step in harvesting is canonicalization—mapping diverse surface names (e.g., "Apple Inc." and "The company from Cupertino") to a single, unique entity ID. To maintain precision, systems employ consistency constraints. By applying logical rules (e.g., "a person can only have one birthplace") and joint inference, systems can resolve contradictions and prune the errors typically found in raw extraction.

The Landscape of Knowledge Bases

The field has been shaped by both academic and industrial efforts. Projects like YAGO (which unified WordNet and Wikipedia) and DBpedia succeeded by first targeting Wikipedia’s semi-structured elements. These academic foundations paved the way for massive industrial knowledge graphs, such as the Google Knowledge Graph and Microsoft Satori, which now underpin global search infrastructure.

Key Achievements

Knowledge harvesting has achieved significant milestones:

  • Scale: Systems have scaled from thousands of hand-crafted entries to billions of relational facts.
  • Taxonomic Depth: Projects like YAGO have successfully constructed high-quality taxonomies with over 300,000 fine-grained classes.
  • Practical Impact: These methods powered IBM Watson’s Jeopardy! win and are now central to commercial AI and search engines.

Future Frontiers and Challenges

Despite these successes, several gaps remain for future research.

Beyond Binary Triples

While the Subject-Predicate-Object (SPO) model is the industry standard, it struggles with "higher-arity" relations—complex events or temporal dynamics that involve multiple entities and timeframes simultaneously.

Temporal and complex relations for Steve Jobs, illustrating the need for data models that capture time intervals.

The Long Tail and Commonsense

Current systems excel at harvesting facts about celebrities but struggle with "long-tail" entities like regional politicians. Furthermore, computers still lack "commonsense knowledge"—the physical properties of objects and the typical flow of human activities that humans take for granted.

Socio-Cultural Context

Finally, facts are not always objective. The "inventor" of a technology may vary by cultural perspective. Future KBs must evolve to capture these socio-cultural nuances rather than assuming a single, universal truth.

Summary updated

Abstract

This article gives an overview on knowledge harvesting: automatically constructing large high-quality knowledge bases from Internet sources. The first part reviews key principles and best-practice methods. The second part points out open challenges for future research.
BibTeX
@inproceedings{weikum2019knowledge,
  author = {Gerhard Weikum and Johannes Hoffart and Fabian M. Suchanek},
  title = {{Knowledge Harvesting: Achievements and Challenges}},
  booktitle = {Computing and Software Science},
  year = {2019},
  url = {https://openreview.net/forum?id=xGwgRvEEWAc},
}
Figure