Modelling epistemic uncertainty in IR evaluation

Modern information retrieval (IR) test collections violate the completeness assumption of the Cranfield paradigm. In order to maximise the available resources, only a sample of documents (i.e. the pool) are judged for relevance by a human assessor(s). The subsequent evaluation protocol does not make any distinctions between assessed or unassessed documents, as documents that are not in the pool are assumed to be not relevant for the topic. This is beneficial from a practical point of view, as the relative performance can be compared with confidence if the experimental conditions are fair for all systems. However, given the incompleteness of relevance assessments, two forms of uncertainty emerge during evaluation. The first is Aleatory uncertainty, which refers to variation in system performance across the topic set, which is often addressed through the use of statistical significance tests. The second form of uncertainty is Epistemic, which refers to the amount of knowledge (or ignorance) we have about the estimate of a system's performance.