Inter-Annotator Agreement Networks

This work develops a simple information theoretic framework that captures the dynamic of the inter-annotator agreement process and unifies a wide range of approaches in unsupervised learning. Our model consists of a pair of annotators whose goal is to maximize the mutual information between their annotations. Training the model with standard stochastic gradient descent is challenging, but we find an ablation of the model that admits variational approximation to be empirically effective. We illustrate the strength our framework by achieving new state-of-the-art accuracy on unsupervised part-of-speech tagging, in particular 78.7% on the 45-tag Penn WSJ dataset. We also show clear performance improvement in unsupervised entity typing.