data annotation
/ DAY-tuh an-oh-TAY-shun /
Supervised learning needs answers to learn from, and those answers rarely come for free — usually a person has to attach them. Data annotation (or labeling) is the work of marking up raw data with the correct answers: drawing a box around each pedestrian in a photo, tagging a tweet as positive or negative, transcribing a recording, marking a tumor in a scan. It is the human-supplied truth a model imitates.
Annotation ranges from quick to painstaking. A simple task is one click per item (spam or not); a hard one might require a radiologist to outline organs pixel by pixel. Because the labels become the model's definition of "correct," their quality is everything. Annotators disagree, get tired, and misread guidelines, so serious projects measure inter-annotator agreement (do independent labelers say the same thing?), write careful instructions, and have multiple people label the same items to catch and resolve conflicts.
Why it matters: labels are often the most expensive, slowest, and most quietly consequential part of a machine-learning project. If the labels are noisy or biased, the model faithfully learns that noise or bias — there is no algorithm that fixes wrong answers it was told are right. This is why so much effort now goes into reducing labeling cost (active learning, where the model asks for labels only on the examples it finds most confusing) and into auditing label quality, the heart of data-centric AI.
To train a self-driving car's vision system, annotators draw tight boxes around every car, cyclist, and pedestrian in hundreds of thousands of street photos. Two people label each image; where they disagree on a hard, half-hidden cyclist, a third reviewer decides — and that careful agreement becomes the truth the model learns.
Two labelers, one referee — manufacturing the 'truth' a model will imitate.
Labels are ground truth only by convention, not by nature. They embody the judgments, fatigue, and biases of the people who made them — so a model can never be more reliable than the labels it was trained on.