Classification & NER

Labeling whole texts and spans within them.

L2 · IntermediateEvolvingKnown~12 min

Reviewed by YuvarajPublished 25 Aug 2026

Text is messy, but two tasks tame most of it. Text classification reads a whole span of text and assigns it a single label, is this review positive or negative, is this email spam, does this support ticket belong to "billing" or "shipping"? Named Entity Recognition (NER) works one token at a time, tagging which words name a person, an organization, or a place. The first answers "what is this document about?"; the second answers "what specific things are named inside it?" Almost every applied NLP pipeline, search, moderation, analytics, knowledge extraction, is assembled from these two shapes.

Two shapes, two label schemes

Text classification

One label (or a small set) for the entire input
Sentiment, spam detection, topic and intent routing
Output is a single class from a fixed set
Scored with document-level accuracy, precision, recall, F1
The model pools the whole text into one representation

Sequence labeling (NER)

One label per token, in reading order
NER, part-of-speech tagging, chunking
Output is a label sequence the same length as the input

Ask the tutor

Ask about this lesson, or about anything in AI. Answers cite the lessons they draw on.

Finished this lesson?

Mark it complete to earn XP, keep your streak, and schedule a review.

Scored at the entity (span) level, not per token

The model must respect span boundaries across tokens

How labels attach to tokens: BIO tagging

The catch with NER is that entities span multiple tokens ("San Francisco", "Satya Nadella"), and two entities of the same type can sit side by side. A flat PER/ORG/LOC tag per token cannot say where one entity stops and the next begins. BIO tagging fixes this with three prefixes:

B- marks the Beginning token of an entity span
I- marks a token Inside (continuing) the current span
O marks a token Outside any entity

Combined with an entity type you get labels like B-PER, I-PER, B-ORG, B-LOC. A person named across two tokens becomes B-PER I-PER; the next word drops back to O. This turns "find the entities" into a per-token classification problem that a sequence model can learn.

1TokenizeSplit the sentence into tokens; each will receive exactly one tag.
2Predict per tokenThe model emits one BIO plus type label for every token, in order.
3Decode spansWalk the tag sequence: each B- opens a span, a matching I- extends it, and O closes it.

A fully tagged example

Take the sentence "Satya Nadella runs Microsoft from Redmond." Its gold BIO tags:

Token	Tag
Satya	B-PER
Nadella	I-PER
runs	O
Microsoft	B-ORG
from	O
Redmond	B-LOC
.	O

Decoding recovers three entities: Satya Nadella (PER), Microsoft (ORG), and Redmond (LOC). The same sequence written inline:

Satya/B-PER Nadella/I-PER runs/O Microsoft/B-ORG from/O Redmond/B-LOC ./O

Scoring entities, not tokens

NER is judged at the entity level: a predicted span is a true positive only if its boundaries and its type both match the gold span exactly. Predicting B-PER on "Satya" but missing "Nadella" is a boundary error, one false positive and one false negative, with no partial credit. Counting true positives, false positives, and false negatives over spans:

P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}, \qquad F_1 = \frac{2PR}{P + R}

Precision asks how many predicted entities were correct; recall asks how many real entities you found; $F_1$ is their harmonic mean. Token-level accuracy looks deceptively high because most tokens are O, so always report entity-level $F_1$ for NER.

Common mistakes

Reporting token accuracy for NER. With O dominating, a model that predicts almost nothing still scores well, entity-level $F_1$ is the honest metric.
Giving partial credit for near-miss spans. Standard scoring (CoNLL, seqeval) demands an exact boundary and type match.
Mismatched tokenization between training and inference, which shifts every tag and silently breaks span decoding.
Ignoring invalid tag sequences, such as an I-ORG that follows an O or a B-PER; CRF decoders and constrained decoding enforce legal transitions.
Treating one classification threshold as universal, tune the decision boundary per class on imbalanced spam or sentiment data.