Apappu writeup of Klein and Manning

This is a review of the paper Klein_2002_conditional_structure_versus_conditional_estimation_in_nlp_models by user:apappu.

Conditional structure vs Conditionaal estimation on NLP models

parameter estimation vs model structures

shows POS task shows errors are not only from label bias but also because independence assuptions of the conditional model structure

experiments on: POS tagging experiments, WSD

• I liked they way teased apart the problem into analyzing condition parameter estimation and the model structure seperately.

• Authors reiterate the statement that it is advantegeous to maximize the objective functions (of parameters) which are similar to eval criteria.

• Effectiveness of smoothing methods is stated through the improvement in accuracy of NB-SCL over NB-CL.

• It seems that NB-CL leverage on Larger training sets, as it fails to do well on smallest training sizes when compared with NB-JL. (but, NB-CL seems to handle rare cases pretty well)

• It is interesting to note tasks like WSD get good numbers using smaller training corpus

• As, authors discuss about model structures, they discuss about CMM and HMM, one is motivated to maximize CL and another to maximize JL.

• But, CMMs are prone to label bias and observation bias.

• Authors claim that independence assumptions manifested in conditional strucutres cause drop down in accuracy.

• They also put forth that explaining-away in conditional structured models which invariably means causing label bias and observation bias.

• Overall, this paper is pretty comprehensive and diagnostic in nature. Especially, I like it the way it emphasizes on why NLP problems are to be dealt differently from other problems using same learning techniques.