WhenFactsExpire

This dataset is described in the article "When Facts Expire: Benchmarking Temporal Validity in Knowledge Graphs," under review at CIKM'2025. The source code for this resource can be found on the Github repository: WhenFactExpire. Description This benchmark addresses the critical challenge of temporal fact validation in Knowledge Graphs (KGs). While traditional KGs often treat facts as timeless, many are only valid within a specific temporal context (e.g., a person's employment at a company). This resource provides a systematic way to develop and evaluate models on their ability to recognize the temporal plausibility of facts. The dataset is derived from the Wikidata dump of May 2023. It has been transformed to support a temporal validation task by generating negative samples through the corruption of the temporal context of true facts, while the subject, predicate, and object remain unchanged. This process ensures that models are evaluated specifically on their temporal reasoning capabilities. The data is provided in multiple configurations to test different model capabilities, varying by: Temporal Scope: A reduced, dense scope (1900-2023) and a full, sparser scope (-1000-2023). Graph Size: Four sizes from Extra-Small to Large, based on entity connectivity. Temporal Granularity: Year-level and Day-Month-Year level precision. Contents Processed.zip: The complete dataset with positive and negative samples organized by the configurations described above. NaturalLanguageDescription.zip: Contains textual descriptions (where available) for entities, relations, and classes in the dataset. Res_TKBI.zip: Contains the results produced by the algorithms referenced in the associated GitHub repository. WhenFactExpire-main.zip: The source code to generate the data and examples to use the data.