2001 · Discriminative Sequence Labelling
Score the tags given the sentence—not a story about how the sentence was generated.
An HMM models P(words, tags): tags transition, then emit words. A Conditional Random Field models P(tags | words) directly. It can combine overlapping clues without claiming they are independent: the word itself, its suffix, its neighbours, and the previous tag all cast weighted votes.
Every complete tag path receives the sum of its feature weights. Viterbi finds the highest-scoring path. A forward pass sums over every possible path into Z(x), turning that score into a normalized conditional probability.
In “they rose,” the transition feature overwhelms the word's slight noun preference. In “the rose,” a Determiner→Noun feature changes the decision.
Try it — decode the whole sequence
Features are engineered. Real CRFs learn weights, but a person still decides what evidence the model may use.
Labels are expensive. Conditional training needs annotated sequences.
Context remains local. Capturing distant structure requires an ever-growing feature inventory.