Nlao writeup of Sutton 2004

This is a review of Sutton_2004_collective_segmentation_and_labeling_of_distant_entities_in_information_extraction by user:Nlao.

I dont find this paper very impressive. The algorithm is a simplified version of Bunescu and Mooney (2004) and avoids the candidate generation step. The evaluation is only on one data set, and the improvement is significant (3%) only on one sub-task.

Different than RMN, the proposed method adds dependency to each pair of same tokens. This leads to n^2 pairwise dependecies, and bias towards same labels when a token appears many times (as explained by Krishnan, 2006).

But still, this is a new method with positive result.

[minor points]