A label is an indication of category, meaning or destination associated with a data point. In machine learning, labels are the answers the model tries to predict. In natural language processing, labels can indicate the sentiment of a sentence, the topic of a document or the type of entity.
How it works
Data labeling involves: defining categories, annotating each example with the correct label, validating consistency between annotators and managing data quality. High quality labels are essential to train accurate models and reduce bias in the machine learning system.
Practical example
In an email dataset, each email can be labeled as spam, normal or important. A classification model learns to predict which label a new email will receive based on labeled examples. In computer vision, each image can have labels indicating the objects present.
Related terms
- supervised-learning: learning with labeled data
- dataset: collection of labeled data
- data-annotation: labeling process
How to test today
Create a simple dataset, e.g., movie reviews labeled as good or bad, and train a simple classification model using scikit-learn. See how label quality affects accuracy.
Frequently asked questions
- What is a label? It is an indication of category or meaning associated with a data point, used in supervised learning and classification.
- Why do label quality matter? Poor labels lead models to learn wrong patterns, increasing error, algorithmic bias and reducing the reliability of predictions.
- How to ensure labeling quality? Use clear category definitions, annotator training, cross validation of labels and annotator agreement metrics.