Liuliu writeup of Sutton 2004

This is a review of Sutton_2004_collective_segmentation_and_labeling_of_distant_entities_in_information_extraction by user:Liuliu.

This work proposed a new general CRF model named skip-CRF for named entity recognition which captures non-local information. Particularly, this model captures the non-local dependencies that same capitalized tokens tend to have the same label.

The original sequence model which is based on Markov assumption cannot model global dependencies otherwise the model will be very complicated. This work extends the original sequence CRF by adding edges between identical capitalized tokens based on input data. This makes the original inference method intractable so that they use loopy belief propagation for approximate inference.

From the results, we could see the error rates decrease a lot for location and speaker but error increases for time. I think this makes sense:

• (1) The global dependencies they capture is that identical capitalized words, which only occurs a lot for location and speaker but not time. The time recognition doesn't benefit a lot from this method.
• (2) The inference is not an exact inference but an approximate one, which might explain the error rates for time increase.

Anyway, I like this model in the way that it combines segmentation and labeling as one problem but not treat them separately as what RMN model does.