A collaborative approach to IR evaluation

Abstract

In this thesis we investigate two main problems: 1) inferring consensus from disparate inputs to improve quality of crowd contributed data; and 2) developing a reliable crowd-aided IR evaluation framework.

With regard to the first contribution, while many statistical label aggregation methods have been proposed, little comparative benchmarking has occurred in the community making it difficult to determine the state-of-the-art in consensus or to quantify novelty and progress, leaving modern systems to adopt simple control strategies. To aid the progress of statistical consensus and make state-of-the-art methods accessible, we develop a benchmarking framework in SQUARE, an open source shared task framework including benchmark datasets, defined tasks, standard metrics, and reference implementations with empirical results for several popular methods. Through the development of SQUARE we propose a crowd simulation model that emulates real crowd environments to enable rapid and reliable experimentation of collaborative methods with different crowd contributions. We apply the findings of the benchmark to develop reliable crowd contributed test collections for IR evaluation.

As our second contribution, we describe a collaborative model for distributing relevance judging tasks between trusted assessors and crowd judges. Based on prior work's hypothesis of judging disagreements on borderline documents, we train a logistic regression model to predict assessor disagreement, prioritizing judging tasks by expected disagreement. Judgments are generated from different crowd models and intelligently aggregated. Given a priority queue, a judging budget, and a ratio for expert vs. crowd judging costs, critical judging tasks are assigned to trusted assessors with the crowd supplying remaining judgments. Results on two TREC datasets show significant judging burden can be confidently shifted to the crowd, achieving high rank correlation and often at lower cost vs. exclusive use of trusted assessors.