NELL2RDF: Reading the Web, Tracking the Provenance, and Publishing it as Linked Data

Estevam Hruschka, José M. Giménez-García, Maísa Cristina Duarte, Antoine Zimmermann, Christophe Gravier, Estevam R. Hruschka Jr., Pierre MaretCkgsemstats@iswc 2018

Abstract

This paper presents NELL2RDF, a system that transforms the massive output of Carnegie Mellon’s Never-Ending Language Learning (NELL) system into a standardized, Linked Data format. By converting over 14.5 billion triples and including rich provenance and confidence metadata, the authors have made one of the world's largest open knowledge bases accessible and queryable for the broader Semantic Web community. The dataset utilizes the PROV-O ontology to capture the "why" and "how sure" behind every extracted fact, bridging the gap between machine learning outputs and standardized web tools.

Introduction

For over a decade, Carnegie Mellon’s NELL has been "reading the web"—continuously extracting facts about entities and their relationships. While NELL is a goldmine of information, its data was historically trapped in an ad-hoc CSV format that was difficult for outside researchers to use.

Crucially, NELL doesn't just store facts; it stores "beliefs" with varying degrees of confidence and complex provenance (e.g., which specific algorithm extracted the fact and in which iteration). Standard RDF (Resource Description Framework) is designed for simple binary relations (Subject-Predicate-Object), making it difficult to represent this "meta-knowledge" without losing the context that makes NELL valuable.

Methodology

The authors developed a pipeline to convert NELL’s tab-separated beliefs into a structured RDF dataset. Key technical choices include:

  • Ontology Mapping: They utilized the PROV-O (Provenance Ontology) to describe how each belief was generated, by which component, and at what time. This mapping allows NELL's internal logic to be expressed through standardized provenance classes.
The NELL2RDF ontology mapping NELL beliefs and provenance to PROV-O classes.
  • Reification Strategies: Since there is no single "best" way to represent statements about statements in the Semantic Web, the authors published the dataset using five different approaches: Standard RDF Reification, N-Ary Relations, Named Graphs, Singleton Properties, and NdFluents.
  • Efficient Storage: To handle the sheer scale of the data, they used HDT (Header Dictionary Triples), a binary format that allows for high compression and fast querying without decompressing the entire file.

Datasets and Results

The paper focuses on iterations 1075, 1090, and 1100 of the NELL system. The resulting NELL2RDF dataset contains over 14.5 billion triples. Key achievements include:

  • Scale: Successfully scaled the conversion from a previous attempt of 5.8 million triples to a massive 14.5 billion.
  • Interoperability: The authors successfully interlinked 296,255 entities from NELL to DBpedia using Wikipedia page metadata, allowing NELL data to be combined with other major datasets in the Linked Open Data cloud.
  • Standardization: Published the data under a Creative Commons CC0 license at a canonical URL (http://w3id.org/nellrdf), complete with SPARQL endpoints for querying.

Discussion

Metadata as the Main Course

The authors argue that NELL’s real value isn't just the facts, but the metadata. By exposing the "why" and "how sure" behind every fact, they've created a massive testbed for research into data quality and uncertainty. This project acts as a bridge between the Machine Learning community (who built NELL) and the Semantic Web community.

The Cost of Complexity

The study found that while most reification methods result in similar file sizes, the Singleton Property approach is an outlier. It resulted in files 30% larger in HDT format because it generates a massive number of unique properties, which breaks standard compression and indexing efficiencies.

Conclusion and Future Work

NELL2RDF successfully democratizes access to a decade of machine learning progress. Future work will focus on processing NELL’s entire 8-year history to allow researchers to track how "beliefs" evolve over time. Additionally, the authors plan to incorporate multilingual support for NELL's expanding Portuguese and French datasets and improve entity resolution to increase the density of links between NELL and the rest of the Semantic Web.

Summary updated

BibTeX
@inproceedings{hruschka2018nell,
  author = {Estevam Hruschka and José M. Giménez-García and Maísa Cristina Duarte and Antoine Zimmermann and Christophe Gravier and Estevam R. Hruschka Jr. and Pierre Maret},
  title = {{NELL2RDF: Reading the Web, Tracking the Provenance, and Publishing it as Linked Data}},
  booktitle = {Ckgsemstats@iswc},
  year = {2018},
}
Figure