Data Labeling
Data labeling is the process of assigning informative tags, categories, or ground-truth values to raw data so it can be used to train supervised machine learning models.
Definition
Data labeling is the process of assigning informative tags, categories, or ground-truth values to raw data so it can be used to train supervised machine learning models.
Overview
Supervised machine learning models learn to map inputs to outputs by example, which requires a dataset of inputs paired with correct, known outputs — labels. Data labeling is the process of producing those labels, whether by human annotators, automated heuristics, weak supervision techniques, or a combination of methods. The quality, consistency, and volume of labeled data are often the single biggest determinants of a supervised model's performance, more so than model architecture choices in many practical settings. Labeling tasks vary widely by data type and application: image classification requires assigning a category to each image; object detection requires drawing bounding boxes around objects of interest; semantic segmentation requires labeling every pixel; text classification requires assigning categories or sentiment to documents; and named entity recognition requires marking spans of text with entity types. Labeling can be performed in-house, through crowdsourcing platforms like Amazon Mechanical Turk, through specialized data labeling vendors and platforms such as Scale AI or Labelbox, or increasingly with the assistance of pre-trained models that propose labels for humans to verify and correct — a workflow often described as human-in-the-loop labeling. Label quality control is a major practical challenge: common techniques include having multiple annotators label the same example and measuring inter-annotator agreement, maintaining detailed labeling guidelines and rubrics, running gold-standard test examples with known answers to catch annotator drift, and using active learning to prioritize which examples most need human labeling. Data labeling is closely related to but distinct from data annotation, a broader term that also covers richer metadata beyond simple category labels, such as transcriptions, relationships, or structured attributes.
Key Concepts
- Assigns ground-truth labels to raw data for supervised model training
- Varies by task: classification labels, bounding boxes, segmentation masks, entity spans
- Can be performed manually, via crowdsourcing, or with automated/weak supervision
- Human-in-the-loop workflows combine model-generated suggestions with human review
- Label quality is often the single biggest driver of downstream model performance
- Quality control methods include multi-annotator agreement checks and gold-standard tests
- Active learning prioritizes the most informative examples for labeling
- Supported by dedicated labeling platforms such as Labelbox and Scale AI