Liuy 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:Liuy.

The authors use two-stage CRF to model this nonlocal dependency, and empirically prove its advantage than other methods that try to model non-local dependencies. They shows the two-stage approach has advantages over other methods modeling non-local dependencies, in both F-measure and time efficiency.

The improvement Pairwise dependences for different occurance of a same entity complicates the dependency graph inference. This work tries not to enforce pairwise dependencies and allow the label dependent on the summary of other labels belonging to the same entity. The paper's novelty is to use in the second stage, the features extracted from the first stage CRF's output. It set up a model to explore dependencies across documents. This is significantly different from conventional sequence models, that is unable to capture the long distance structure. The features used in the paper are all functions of the output of the first stage CRF. They are token-majority, entity-majority, and superentity-majority.

I like this work mainly because it is easy to implement and is quite efficient, although there are two stages involved in the algorithm. It only takes twice as long as the basic CRF, and thus is quite efficient compared with those methods using approximate inference. The superentity-majority used in the work allows entity sequences to use labels of its supersequence. However, one limitation of this work is that this work only tests the effectiveness of their method on the CoNLL03 English NER. I would like to see more evidence from other language NRE and other NER datasets.