Feature-based structured models provide a flexible and elegant framework for various information extraction (IE) tasks. These include label sequences for traditional IE, segmentation models for entity-level extractions, and skip chain models for collective labeling. I will present efficient inference algorithms for finding the highest scoring (MAP) prediction for two interesting types of structured models in IE.
There are two popular formulations for maximum margin training of structured spaces: margin scaling and slack scaling. While margin scaling is extremely popular since it requires the same kind of MAP inference as prediction, slack scaling is believed to be more accurate and better-behaved. I will describe an efficient variational approximation to the slack scaling method that solves its inference bottleneck while retaining its accuracy advantage over margin scaling. Further I argue that existing scaling approaches do not separate the true labeling comprehensively while generating violating constraints. I will propose a new max-margin trainer PosLearn that generates violators to ensure separation at each position of a decomposable loss function.
Bio