JustEvents: A Crowdsourced Corpus for Event Validation with Strict Temporal Constraints

Abstract

Inspecting text to affirm the occurrence of an event is a non-trivial task. Since events are tied to temporal attributes, this task is more complex than merely identifying evidence of entities acting together and thus defining the event in a document. Manual inspection is a typical solution, although it is an onerous task and becomes infeasible with an increasing scale of documents. Therefore, the task of automatically determining whether an event occurs in a document or corpus, named as event validation, has been recently investigated. In this paper, we present a dataset for benchmarking event validation methods. Events and documents are coupled in pairs, whose validity has been judged by human evaluators based on whether the document in the pair contains evidence of the given event. In contrast to the notion of relevance considered in available datasets for event detection, validity judgments in this work strictly consider whether a document reports an event within its timespan as well as the number of event participants reported in the document. These requirements make the generation of manual validity judgments an onerous procedure. The ground truth, made of multiple judgments for each pair, has been acquired through crowdsourcing.