Stability and Generalization in Structured Prediction

This work proposes new PAC-Bayesian generalization bounds for structured prediction that decrease as a function of both the number of examples and the size of each example, and applies them to several common learning scenarios.