KeisukeKamataki writeup of Krishnan 2006

This is a review of Krishnan_2006_an_effective_two_stage_model_for_exploiting_non_local_dependencies_in_named_entity_recognition by user:KeisukeKamataki.

• Summary:

They used two-stage approach of CRF to solve the non-local dependency problem in named entity recognition. First, they used the standard linear-chain CRF with popular features. Then, they introduced additional features from the output of the first CRF. Specifically, they tried Token-majority, Entity-majority, and Superentity-majorigy features. They also tried a lot of tweeks to train good model such as cross-validations. Their method generally got about 1-2% F-measure improvement from their original model.

• I like:

Their method come from detailed statistical analysis of real token entity data and reasonale. They are also clear about their advantages and how to use in practice.

• I didn't like/well understand

This paper didn't look so exciting to me. This might be because the simplicity to use same algorithm twice makes computational time exactly twice(although intuitivly sounds reasonable approach). Simplicity is usually a good thing in algorithms, but their simplicity just solves very restricted problem sacrificing computational time.