A study of imitation learning methods for semantic role labeling

Abstract

Global features have proven effective in a wide range of structured prediction problems but come with high inference costs. Imitation learning is a common method for training models when exact inference isn't feasible. We study imitation learning for Semantic Role Labeling (SRL) and analyze the effectiveness of the Violation Fixing Perceptron (VFP) (Huang et al., 2012) and Locally Optimal Learning to Search (LOLS) (Chang et al., 2015) frameworks with respect to SRL global features. We describe problems in applying each framework to SRL and evaluate the effectiveness of some solutions. We also show that action ordering, including easy first inference, has a large impact on the quality of greedy global models.