Philgoo Han writeup of 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:Ironfoot.

• Upgrade of non-local feature CRF
  • Soft contraint of consistency among long distance similar tokens
  • No need of approximate inference
• Statistics of label consistency in document level and corpus level
  • Good and detail starting point for raising the issue
• Two step CRF
  • First step: normal CRF
  • Second step: CRF with non-local features derived from first step
    • In general dividing the problem into sequential steps may lead error propogation. But in this case it is error correction. Interesting