Defeating Tyranny of the Masses in Crowdsourcing: Accounting for Low-Skilled and Adversarial Workers

Crowdsourcing has emerged as a useful learning paradigm which allows us to instantly recruit workers on the web to solve large scale problems, such as quick annotation of image, web page, or document databases. Automated inference engines that fuse the answers or opinions from the crowd to make critical decisions are susceptible to unreliable, low-skilled and malicious workers who tend to mislead the system towards inaccurate inferences. We present a probabilistic generative framework to model worker responses for multicategory crowdsourcing tasks based on two novel paradigms. First, we decompose worker reliability into skill level and intention. Second, we introduce a stochastic model for answer generation that plausibly captures the interplay between worker skills, intentions, and task difficulties. This framework allows us to model and estimate a broad range of worker "types". A generalized Expectation Maximization algorithm is presented to jointly estimate the unknown ground truth answers along with worker and task parameters. As supported experimentally, the proposed scheme de-emphasizes answers from low skilled workers and leverages malicious workers to, in fact, improve crowd aggregation. Moreover, our approach is especially advantageous when there is an (a priori unknown) majority of low-skilled and/or malicious workers in the crowd.