Sgardine writesup Sutton 2004

This is a review of sutton_2004_collective_segmentation_and_labeling_of_distant_entities_in_information_extraction by user:Sgardine.

Summary

In NER, repeated mentions of the same entity should usually be classified the same. Traditional sequence tagging approaches do not allow information from the classification of one mention to influence others. The authors introduce skip-chain CRFs, in which additional links between similar or identical entities are introduced. Because the resulting graph is no longer a tree, exact inference is intractable, and so loopy BP is used for approximate inference. The model was evaluated on seminar announcements; it was found better than plain CRF on speakers and slightly worse on all other fields (presumably the approximate inference effect overpowered any gains in consistency)

Commentary

I would have liked some attempt to separate the effects of approximate inference (which probably degrade performace) from those of enforcing consistency (which probably improve performace). Maybe training a skip-chain CRF with random links (which would not be expected to help appreciably)?

• That's an interesting experiment to consider... Wcohen