Philgoo Han writeup of Klein and Manning

This is a review of Klein_2002_conditional_structure_versus_conditional_estimation_in_nlp_models by user:Ironfoot.

• Where is the perfomance comming from, the structure, objective function or the feature selection.

• Different objective functions (loss function)
  • Joint likelihood
  • Conditional likelihood: This must be something like logistic regression?
  • Sum of conditional likelihood
  • Any other important objectives?
• Scoring
  • Accurcacy : Why use this specific scoring?
• Optimization
  • What is a deficient model? Probability may not sum up to 1?
  • How do you know a certain local optima is global in actual?
• Result
  • CL is generally better than JL
    • Any tunning possible for rare observations in JL?
    • Can we make a criterion depending on the data(feature) without actually testing both?

• Different model structure
  • HMM
    • Objective function: JL, CL
  • CMM
    • Observation bias
    • Label bias is not much likely to occur