sequence labeling
/ SEE-kwens LAY-bul-ing /
Sequence labeling is the family of tasks where you walk through a sequence — usually the words of a sentence — and attach a label to each item in turn, never losing track of the order. Picture a teacher reading a sentence aloud and, word by word, calling out a category for each one. It's different from sorting a whole document into one bucket; here every single position gets its own answer, and each answer should fit its neighbors.
Many famous NLP jobs are secretly the same shape. Part-of-speech tagging labels each word with its grammar role. Named-entity recognition labels each word as the start of a name, inside a name, or not a name at all. Even chopping unspaced text into words can be cast this way. The thread connecting them: a label for every token, and the labels are not independent — "inside an organization name" only makes sense if the previous word was "start of an organization name," so good models reason about whole label sequences, not one position in a vacuum.
This is worth naming because seeing the shared structure pays off. Classic approaches used models built to respect label dependencies; modern ones use neural networks (often the kind that read both left and right) that produce a label per token. The honest reminder is the same as for its members: these systems learn from text humans painstakingly labeled, they lean heavily on context, and they wobble on informal language, unfamiliar words, and languages with little annotated data.
For NER on "Steve Jobs founded Apple," the label sequence might be: Steve = B-PERSON (begin), Jobs = I-PERSON (inside), founded = O (outside), Apple = B-ORG. "Inside-a-person" is only valid right after "begin-a-person" — the labels constrain each other.
One label per token, and the labels constrain one another — that's the sequence-labeling shape.
The defining twist is that labels aren't independent — each one depends on its neighbors, so the model predicts a coherent whole sequence, not isolated guesses. POS tagging and NER are both just sequence labeling in disguise.