Sequence Labelling

Give every token a tag: this word starts a person's name, this one continues it, this one is outside any entity. It is the oldest shape of structured prediction still in daily use, and the reason it is not just per-token classification is that the tags constrain each other — a continuation tag with nothing to continue is not a mistake about one token, it is an invalid structure.

The BIO scheme encodes that: B for the beginning of an entity, I for inside, O for outside, with I-PER only legal after B-PER or I-PER. Evaluation is at the entity level rather than the token level, which is the detail that changes what a model is optimising for — an entity counts only if every one of its tokens is right and its boundaries are exact, so partial credit does not exist.

That all-or-nothing scoring makes the entity-level number fall much faster than token accuracy suggests. An entity of length L survives only if all L tokens are correct, so at a per-token error rate e the share of intact entities goes as (1−e) super ^L  — a 2% token error rate leaves 98% of single-token entities but only about 90% of five-token ones. Long entities are where a model looks much worse under the metric people actually report.