The Effect of Document Order and Topic Difficulty on Assessor Agreement

Abstract

Human relevance judgments are a key component for measuring the effectiveness of information retrieval systems using test collections. Since relevance is not an absolute concept, human assessors can disagree on particular topic-document pairs for a variety of reasons. In this work we investigate the effect that document presentation order has on inter-rater agreement, comparing two presentation ordering approaches similar to those used in IR evaluation campaigns: decreasing relevance order and document identifier order. We make a further distinction between "easy" topics and "hard" topics in order to explore system effects on inter-rater agreement. The results of our pilot user study indicate that assessor agreement is higher when documents are judged in document identifier order. In addition, there is higher overall agreement on easy topics than on hard topics.