Provenance Information in a Collaborative Knowledge Graph: An Evaluation of Wikidata External References

Wikidata is a collaboratively-edited knowledge graph; it expresses knowledge in the form of subject-property-value triples, which can be enhanced with references to add provenance information. Understanding the quality of Wikidata is key to its widespread adoption as a knowledge resource. We analyse one aspect of Wikidata quality, provenance, in terms of relevance and authoritativeness of its external references. We follow a two-staged approach. First, we perform a crowdsourced evaluation of references. Second, we use the judgements collected in the first stage to train a machine learning model to predict reference quality on a large-scale. The features chosen for the models were related to reference editing and the semantics of the triples they referred to. 61% of the references evaluated were relevant and authoritative. Bad references were often links that changed and either stopped working or pointed to other pages. The machine learning models outperformed the baseline and were able to accurately predict non-relevant and nonauthoritative references. Further work should focus on implementing our approach in Wikidata to help editors find bad references.