Incremental Quality Inference in Crowdsourcing

Abstract

Crowdsourcing has attracted significant attention from the database community in recent years and several crowdsourced databases have been proposed to incorporate human power into traditional database systems. One big issue in crowdsourcing is to achieve high quality because workers may return incorrect answers. A typical solution to address this problem is to assign each question to multiple workers and combine workers’ answers to generate the final result. One big challenge arising in this strategy is to infer worker’s quality. Existing methods usually assume each worker has a fixed quality and compute the quality using qualification tests or historical performance. However these methods cannot accurately estimate a worker’s quality. To address this problem, we propose a worker model and devise an incremental inference strategy to accurately compute the workers’ quality. We also propose a question model and develop two efficient strategies to combine the worker’s model to compute the question’s result. We implement our method and compare with existing inference approaches on real crowdsourcing platforms using real-world datasets, and the experiments indicate that our method achieves high accuracy and outperforms existing approaches.