Text annotation is the labeling of text data so a natural language processing model can learn from it, by tagging entities, classifying documents, marking sentiment, or linking spans. It is the NLP counterpart to image annotation, applied to words rather than pixels.
Text annotation attaches labels to written language so models can understand it. The labels operate at different levels: a whole document (spam or not, a topic, a sentiment), a span of text (this phrase is a person, this is a date, named-entity recognition), the relationship between spans (this pronoun refers to that name), or token-level tags (part of speech).
It is the supervised foundation of natural language processing, the same role image annotation plays for vision. Increasingly, text annotation also includes labeling model outputs, ranking responses or flagging errors, the human feedback behind instruction-tuned and aligned language models.
Key takeaways
Text annotation labels written language for NLP, at the document, span, relationship, or token level.
Common tasks include classification, named-entity recognition, sentiment, and coreference.
It now also covers labeling LLM outputs, the human feedback behind aligned models.
What text annotation provides
Common text annotation tasks.
Common text annotation tasks.
Task
What it labels
Classification
Topic, intent, sentiment, spam
Named-entity recognition (NER)
Tag spans as people, places, dates, and more
Span and relationship labeling
Link mentions, mark coreference
Token-level tagging
Part of speech, segmentation
Output labeling
Rating or ranking model responses, for RLHF and evaluation
How it works
Annotators work at the level the task requires, tagging whole documents, marking spans of text as entities, linking mentions across a passage, or applying token-level tags, and, for model alignment, rating or ranking generated responses. Clear guidelines matter most for the subjective tasks, where reasonable people read the same text differently.
Why it matters
Text annotation is the backbone of NLP and, increasingly, of how large language models are aligned and evaluated. Text's hardest annotation problem is subjectivity. Sentiment, toxicity, intent, and relevance are genuinely contested, so agreement is often inherently lower than in vision, and forcing a single "correct" label hides real disagreement the model then cannot learn to represent. Mature text annotation treats subjective tasks as distributions, capturing multiple annotators' judgments rather than collapsing them, the opposite instinct from the crisp boxes-and-classes mindset of much visual annotation.
Frequently asked questions
What is text annotation?
Labeling written language for NLP, classification, entity tagging, sentiment, and more.
What is named-entity recognition?
A text annotation task that tags spans of text as entities like people, places, and dates.
How does text annotation relate to LLMs?
It includes labeling and ranking model outputs, the human feedback used to align and evaluate language models.