Contextual entity disambiguation in domains with weak identity criteria: Disambiguating golden age amsterdamers

Entity disambiguation is a widely investigated topic, and many matching algorithms have been proposed. However, this task has not yet been satisfactorily addressed when the domain of interest provides poor or incomplete data with little discriminating power. In these cases, the use of content fields such as name and date is not enough and the simple use of relations with other entities is not of much help when these related entities also need disambiguation before they can be used. Therefore, we propose an approach for the disambiguation of clustered resources using context (related entities that are also clustered) as evidence for reconciling matched entities. We test the proposed method on datasets of historical records from Amsterdam in the 17th century for which context is available, and we compare the results of the proposed approach to a gold standard generated by three experts, which we make available online. The results show that the proposed approach manages to meaningfully use context for isolating identity sub-clusters with higher quality by eliminating potentially false positive matches.