Conditional Random Fields
• Supervised graphical model that blends sequence
labeling with maximum entropy classification
–  A reasonable first baseline to improve upon.
–  SVMs also a reasonable approach but CRFs do sequence
labeling, not classification.
–  Classification done per token, over sentences.
–  A “Not applicable” tag applies to all tokens not of interest.
• We use the implementation from CRF++ (Kudo, 05)
–  Easy to use interface.
–  No feature selection* (still can be improved).