Selen writeup of Sutton 2004

This is a review of Sutton_2004_collective_segmentation_and_labeling_of_distant_entities_in_information_extraction by user:Selen

This paper relates similar entities considering non-local dependencies by introducing skip chain CRF's. Although the idea of adding edges in CRF's to represent dependencies is not a bad idea this paper suffers from two major problems:

Exact inference is intractable so they use approximate inference however, since even with approximate inference they cannot relate all possible related entities together they only consider capitalized words. This means that their method has a very limited chance of generalization.

Sparsity in the graph can be introduced by other means, perhaps they could have considered using a l1 shrinkage before inference to prune some of the edges.