OEKG: The Open Event Knowledge Graph

Simon Gottschalk, Endri Kacupaj, Sara Abdollahi, Diego Alves, Gabriel Amaral, Elisavet Koutsiana, Tin Kuculo, Daniela Major, Caio Mello, Gullal S. Cheema, Abdul Sittar, Swati, Golsa Tahmasebzadeh, Gaurish ThakkarCLEOPATRA @ WWW 2021

Introduction

In an increasingly globalized world, major incidents—from political elections to international sporting events—generate vast amounts of data across different languages and media formats. However, this information is often siloed. News archives, image databases, and specialized knowledge graphs rarely "speak" to one another, making it difficult for researchers to conduct cross-lingual event analytics. The Open Event Knowledge Graph (OEKG) addresses this fragmentation by unifying over 400 million facts into a single, queryable framework. By centering the "event" as the primary entity, OEKG enables a holistic view of how global incidents are perceived and reported across diverse cultures.

Methodology: A Unified Integration Pipeline

The OEKG is built upon EventKG-light, a streamlined version of a pre-existing event-centric knowledge graph. To integrate disparate datasets, the authors developed a robust three-step pipeline:

  1. Entity Linking: New data is mapped to OEKG resource URIs. A web API is provided to translate common Wikidata or DBpedia IDs into OEKG identifiers, lowering the barrier for external researchers.
  2. Graph Creation: Data is transformed into RDF triples using standard vocabularies like schema.org and onyx (for sentiment), ensuring broad interoperability.
  3. Named Graphs: Datasets are stored as "named graphs," allowing them to be queried as a single unit while preserving their original provenance.

A Multi-faceted Data Ecosystem

The OEKG integrates seven distinct datasets to provide a comprehensive view of global events. These include EventKG-light (the backbone with 434M triples), EventKG+Click (user interaction data), VQuAnDa (question-answering pairs), and MLM (multimodal location data).

Multilingual and multimodal representation of events (Notre-Dame fire) and entities (Hanover) in OEKG.

This integration allows for a rich, multilingual representation of entities and events, as seen in the examples of the Notre-Dame fire and the city of Hanover, where labels are provided in multiple languages alongside geographic and visual data.

Key Insights: Events as First-Class Citizens

The core innovation of OEKG is treating the event as the primary join key. This shift enables complex temporal and cross-modal queries that were previously impossible.

  • Visual Event History: By linking images to event locations, researchers can retrieve visual evidence of regions specifically during or after significant incidents.
  • Sentiment Analysis: Integrating sentiment scores directly into the schema allows for tracking the emotional response of different linguistic communities to the same event.
  • Hybrid Question Answering: Linking QA pairs to news articles allows systems to provide both direct answers and the broader journalistic context.
Integration of news articles and question-answering pairs within the OEKG framework.

Results and Utility

The resulting graph contains over 436 million triples. The authors demonstrated its utility through several use cases:

  1. Image Retrieval: Querying for images of earthquake-affected locations using the UNER class hierarchy.
  2. Hybrid QA: Connecting questions about professional wrestling to Spanish news articles, providing both factual answers and media sentiment.
  3. Event Recommendation: Generating Russian-perspective timelines of World War I by filtering for high relevance scores in Russian-language interaction data.

Conclusion and Future Work

While OEKG provides a powerful unified resource, it currently prioritizes ease of integration over exhaustive provenance. Future work will focus on benchmarking the graph's impact on downstream tasks like automated fact-checking and machine translation. The quality of the graph remains dependent on the accuracy of entity linking, an area for ongoing refinement in natural language processing.

Summary updated

BibTeX
@inproceedings{gottschalk2021oekg,
  author = {Simon Gottschalk and Endri Kacupaj and Sara Abdollahi and Diego Alves and Gabriel Amaral and Elisavet Koutsiana and Tin Kuculo and Daniela Major and Caio Mello and Gullal S. Cheema and Abdul Sittar and Swati and Golsa Tahmasebzadeh and Gaurish Thakkar},
  title = {{OEKG: The Open Event Knowledge Graph}},
  booktitle = {CLEOPATRA @ WWW},
  year = {2021},
}
Figure