Conditional random fields model the probability of some hidden states sequence conditioned on the observation sequence. Unlike HMM, which models the joint distribution of hidden states and observation, CRF does not try to model the observation. Based on the maxent principle, CRF model is a product of some exponentials of a weighted combination of some feature vectors. The feature vector only depends on previous state, current state, current observation and current position -- for simplicity. Forward-backward algorithm is used to estimate the weights.
Reviewed by
zzb3886
- 2009-02-25 02:37:28