Learning from Multiple Annotators: When Data is Hard and Annotators are Unreliable

Abstract

The crowd sourcing services became popular making it easy and fast to label datasets by multiple annotators in order to achieve supervised learning tasks. Unfortunately, in this context, annotators are not reliable as they may have different levels of experience or knowledge. Furthermore, the data to be labeled may also vary in their level of difficulty. How do we deal with hard data to label and unreliable annotators? In this paper, we present a probabilistic model to learn from multiple naive annotators, considering that annotators may decline to label an instance when they are unsure. Both errors and ignorance of annotators are integrated separately into the proposed Bayesian model. Experiments on several datasets show that our method achieves superior performance compared to other efficient learning algorithms.