Apappu writeup on Krishnan and Manning

This is a review of krishnan_2006_an_effective_two_stage_model_for_exploiting_non_local_dependencies_in_named_entity_recognition by user:Apappu.

• Authors propose a two-stage CRF model to capture non-local features to improve corpus wide named entity recognition.

• An initial baseline with standard feature classes has been trained using CRFs and on top of it a second-stage CRF is incorporated to address document and corpus level features.

• Second-stage features are motivated by corpus statistics provided in Table 1 and 2, where off-diagonal numbers are less dense.

• I like their two-stage approach since they are doing fine with time taken for inference compared to other approaches on a similar task.

• But, they never describe what do they actually do to enforce soft or hard constraint especially in case of LOC vs ORG labeling.

• They have a very good baseline but what if they have a mediocre baseline and they had to improve on top of it. Also, I don't see corpus level features contributing to the performance when compared to document level features.