Margin-based active learning for structured output spaces

In many complex machine learning applications there is a need to learn multiple interdependent output variables, where knowledge of these interdependencies can be exploited to improve the global performance. Typically, these structured output scenarios are also characterized by a high cost associated with obtaining supervised training data, motivating the study of active learning for these situations. Starting with active learning approaches for multiclass classification, we first design querying functions for selecting entire structured instances, exploring the tradeoff between selecting instances based on a global margin or a combination of the margin of local classifiers. We then look at the setting where subcomponents of the structured instance can be queried independently and examine the benefit of incorporating structural information in such scenarios. Empirical results on both synthetic data and the semantic role labeling task demonstrate a significant reduction in the need for supervised training data when using the proposed methods.