Source-linked AI summary
EventKG: A Multilingual Event-Centric Temporal Knowledge Graph
Simon Gottschalk, Elena Demidova
TL;DR
Event-centric information on the Web, in news, and social media is fragmented across sources, while existing knowledge graphs provide limited event and temporal coverage. EventKG addresses this gap by integrating multilingual event representations and temporal relations into a canonical knowledge graph. It contains over 690 thousand events and over 2.3 million temporal relations, while its coverage remains bounded by limited multilingual descriptions and an explicit lack of distinction among some event types.
Problem
Existing knowledge graphs focus mostly on entity-centric information and provide insufficient coverage and completeness for events and temporal relations needed for semantic analytics.
Method
EventKG extracts and integrates multilingual event representations and temporal relations from large-scale knowledge graphs and semi-structured sources using a canonical representation.
Results
Over 690 thousand events and over 2.3 million temporal relations are included in EventKG V1.1, with light-weight integration increasing coverage and completeness, particularly for times and locations.
Takeaways & Limitations
EventKG provides a multilingual event-centric resource for analytics and applications involving historical and contemporary events, including timeline generation and cross-source event analysis.
Takeaways & Limitations
EventKG V1.1 does not explicitly distinguish single events, seasons, and event series.
Abstract
from arXiv · showhide
One of the key requirements to facilitate semantic analytics of information regarding contemporary and historical events on the Web, in the news and in social media is the availability of reference knowledge repositories containing comprehensive representations of events and temporal relations. Existing knowledge graphs, with popular examples including DBpedia, YAGO and Wikidata, focus mostly on entity-centric information and are insufficient in terms of their coverage and completeness with respect to events and temporal relations. EventKG presented in this paper is a multilingual event-centric temporal knowledge graph that addresses this gap. EventKG incorporates over 690 thousand contemporary and historical events and over 2.3 million temporal relations extracted from several large-scale knowledge graphs and semi-structured sources and makes them available through a canonical representation.
1 Introduction
EventKG addresses the fragmented and incomplete representation of events and temporal relations by integrating multilingual information from heterogeneous knowledge graphs and semi-structured sources into a reusable event-centric resource.
- Existing knowledge graphs primarily provide entity-centric information, leaving event representations incomplete, ambiguously identified, and largely limited to named events and encyclopedic knowledge.
- EventKG integrates event-centric information and temporal relations from heterogeneous knowledge graphs and manually curated semi-structured sources.
- The knowledge graph uses a canonical representation, light-weight integration, relation-strength and popularity indicators, and provenance for its information.
- EventKG covers English, German, French, Russian, and Portuguese and provides an open-source framework extensible to further languages and reference sources.
- Over 690 thousand events and over 2.3 million temporal relations are included in EventKG V1.1.
2 Relevance
EventKG supports multilingual event analytics by combining complementary information across sources, enabling richer timelines, cross-cultural comparisons, and reuse in Semantic Web, NLP, and Digital Humanities applications.
- EventKG can support event-related Question Answering by providing information beyond the entity-centric knowledge graphs commonly used in current research.
- EventKG can generate more complete timelines by combining complementary event information from multiple sources.
- The World War II timeline includes an event extracted from an English Wikipedia event list that is absent from the reference knowledge graphs.
- Event popularity and relation strength vary across language editions and can guide the selection of relevant timeline entries under layout constraints.
- EventKG's use of RDF, SPARQL, open data, and established vocabularies supports reuse across Semantic Web, NLP, and Digital Humanities applications.
3 EventKG Data Model
EventKG provides a lightweight, canonical model for integrating heterogeneous event representations, temporal relations, relation characteristics, and provenance. It builds on SEM while extending it to represent event-centric and entity-centric information across diverse temporal and event relations.
- Data-model goals: EventKG integrates heterogeneous event representations and temporal relations through a lightweight canonical data model with provenance.The model is designed to support event relations, relation descriptions, and source traceability.
- Schema foundation: SEM supplies the event-centric foundation, while EventKG adds classes and properties for temporal relations, event relations, and provenance.EventKG uses SEM as a generic framework and extends it within its own schema.
- Events and entities: Events, places, and participating entities are represented as sem:Event, sem:Place, and sem:Actor, with temporal, spatial, label, alias, and description information.These classes are subclasses of sem:Core, and events connect to locations through sem:hasPlace.
- Temporal relations: Temporal relations cover event-entity, entity-event, and entity-entity connections, including relations whose validity is inferred indirectly from existence times.The model also addresses cases that SEM cannot represent when an entity acts as a subject.
- Event relations: EventKG represents sub-events, event sequences, previous and next events, and location hierarchies using reused vocabulary properties.These relations support event series, seasons, and topic-related event groupings.
- Quantification and provenance: Relation strength and event popularity are quantified using links and mentions, while named graphs preserve provenance for resources, sources, and statements.The links factor counts article references, mentions count sentences mentioning both relation endpoints, and named graphs support source-specific retrieval.
4 EventKG Generation Pipeline
The EventKG pipeline identifies events and relations across knowledge graphs and semi-structured sources, then integrates, fuses, and exports them in the EventKG model. It combines direct identification with recall-oriented heuristics and source-aware processing.
- Pipeline overview: The pipeline collects reference-source dumps, identifies and extracts events, extracts relations, integrates equivalent instances, fuses information, and outputs RDF.Separate named graphs preserve source-specific information alongside the fused EventKG graph.
- Event identification: Events are identified from Wikidata, DBpedia, YAGO, Wikipedia event lists, and WCEP using source-specific classes, categories, temporal titles, and interfaces.YAGO ontology categories are avoided for initial identification because of noisy event subcategories.
- Event identification: 98% average precision was achieved when manually evaluating 100 events per knowledge graph and language edition in DBpedia and Wikidata.This evaluation concerns the events identified in the initial extraction step.
- Recall-oriented heuristics: 94% and 88% precision were obtained for the category-based heuristic on 100 English- and Russian-linked events, respectively.The heuristic uses language-dependent Wikipedia category-name patterns to increase recall.
- Relation extraction: Relations are extracted from temporal validity information, indirect temporal signals, and mappings for event, entity, and sub-event predicates.Temporal relations come from YAGO and Wikidata because DBpedia does not provide the required validity information.
- Integration and fusion: Integration uses owl:sameAs links to combine equivalent real-world instances, while rule-based fusion aggregates their temporal, spatial, and type information.Location fusion uses containment relations to reduce locations, and type fusion follows the DBpedia ontology.
5 EventKG Characteristics
EventKG V1.1 integrates multilingual event and temporal information from knowledge graphs and semi-structured sources, providing broad coverage, substantial temporal relations, and more complete event representations. Its statistics also reveal remaining gaps in multilingual descriptions and spatial information.
- Dataset characteristics: EventKG V1.1 contains over 690 thousand events and over 2.3 million temporal relations extracted from knowledge graphs and semi-structured sources.Nearly half of the events originate from existing knowledge graphs, while 53.25% come from semi-structured sources.
- Dataset characteristics: 76.21% of events in the eventKG-g:event kg graph possess a known start or end time, while locations are provided for 12.21%.The paper identifies location coverage as an area for future enrichment.
- Comparison with reference sources: 690,247 events were identified in EventKG, exceeding the number identified in each reference source through their integration.The paper attributes this especially to combining knowledge graphs with semi-structured sources.
- Comparison with reference sources: EventKG provides more complete event representations than the reference knowledge graphs, including higher coverage of temporal and spatial information.For example, 322,669 EventKG events map to Wikidata, compared with 266,198 events initially identified there.
- Relation and fusion statistics: 93.62% of temporal relations have a start time from 1900 to 2020, and 81.75% of events extracted from knowledge graphs are covered by multiple sources.Among events with known start times, 93.79% agree on their start times across sources.
- Textual descriptions: EventKG contains information in five languages, but only 4.49% of events provide labels in all languages.Among 367,578 events from semi-structured sources, only 115 have descriptions in all five languages.
6 Reusability Aspects
EventKG is designed for reuse through open standards, downloadable and queryable access, extensible RDF modeling, persistent identifiers, and substantial interlinking with reference sources. Its intended applications include event-centric question answering and Web-archive enrichment.
- Access and reuse: EventKG is available for download and through a SPARQL endpoint, with documentation and example queries provided on its homepage.The resource is modeled in RDF and includes substantial interlinking with its reference sources.
- Applications: Intended uses include enriching Web archives with event-centric data and supporting event-centric Question Answering applications.These applications are pursued in collaboration with EU projects such as ALEXANDRIA and WDAqua ITN.
- Open and interoperable publication: EventKG uses RDF, multilingual data, dereferenceable URIs, and persistent URI reuse across versions to support interoperable reuse.The same URIs are consistently reused for the same real-world objects.
- Data modeling: EventKG reuses and extends the SEM event model and existing vocabularies, provides metadata through VoID, and follows FAIR principles.Its description is available in human- and machine-readable formats.
7 Availability & Sustainability
EventKG is publicly available under an open license, and its extraction pipeline is open source. Its sustainability plan combines reusable software, integration of publicly available data, and extensibility to additional sources, languages, and properties.
- Availability: EventKG is publicly available under a persistent URI and the CC BY 4.0 license.The resource uses open standards and provides information on citing it.
- Availability and sustainability: The EventKG extraction pipeline is available as open source software on GitHub under the MIT License.This supports reuse of the software for creating or extending EventKG versions.
- Sustainability: EventKG’s sustainability plan includes reusable open-source architecture and software for adding reference sources, languages, and event properties.The plan also relies on integrating existing publicly available data.
8 Related Work
Related work provides event models, entity-focused knowledge-graph construction, and news-based event extraction, but EventKG concentrates specifically on integrating event-centric information and temporal relations across heterogeneous sources.
- Event models and vocabularies: Existing event models and vocabularies support representing events and linking them with actors, times, and places.Examples include ECKG, SEM, schema.org, and LODE.
- Event-centric knowledge-graph construction: Most automatic knowledge-graph construction and integration approaches focus on entities and related facts rather than events.EventKG instead focuses on events and temporal relations and incorporates events extracted from WCEP.
- News-based extraction: News-based approaches address knowledge-graph construction from plain text and named-event extraction, but automatic news extraction can produce noisy and unreliable results.These methods apply Open Information Extraction and develop it for event-extraction challenges in news.
9 Conclusion
EventKG integrates and harmonizes multilingual event-centric and temporal information for historical and contemporary events. Its lightweight integration improves event coverage and completeness, especially for times and locations, while adding relation-strength, popularity, and provenance information.
- EventKG V1.1 includes over 690 thousand event resources and over 2.3 million temporal relations.
- The knowledge graph integrates and fuses structured and semi-structured multilingual event representations and temporal relations in one graph.
- EventKG provides information for assessing relation strength and event popularity while preserving provenance.
- Lightweight integration significantly increases the coverage and completeness of event representations, particularly for times and locations.