Rbosaghz review of Sutton 2004

This is a review of sutton_2004_collective_segmentation_and_labeling_of_distant_entities_in_information_extraction by user:Rbosaghz.

This paper proposes a new general CRF model (called Skip-CRFs) for the NER task, to capture non-local information. An example of a particular non-local dependency that is modeled with these skip-CRFs is that tokens with the same capitalization pattern often have the same label.

By adding edges to the regular CRF model, the original inference problem becomes intractable. They authors get around this by using loopy belief propagation for approximate inference.